The Machine Objects class library (in short Macho) supports a subset of the UML statechart notation for implementing hierarchical state machines in straight C++, similar in spirit to the GoF "State" design pattern. The currently supported features are hierarchical states, entry and exit actions, state histories and state variables.
For license information please see the MIT License or the header in file Macho.hpp.
| Homepage is at | http://ehiti.sdf-eu.org/macho |
| Feedback goes to | macho@gmx.org |
Copyright 2005 by Eduard Hiti.
You are encouraged to provide any changes, extensions and corrections for this software to the author for inclusion into future versions.
I like the GoF "State" design pattern. It enables implementing the important concept of state machines with common programming language features. By utilising only basic language mechanisms it is easy to apply in real-life software development.
Another important property that stems from this simplicity is orthogonality, meaning that the pattern can be combined with other design elements, patterns and idioms in arbitrary ways.
In contrast stand the tool supported approaches to state machine creation (of which there is no shortage). Based on code generators and special editors, they mostly generate incomprehensible code and forfeit orthogonality by necessarily being outside the domain of the programming language.
Unfortunately the "State" pattern is limited in scope because it does not allow for hierarchical state machines. This is regrettable because flat state machines tend to become unwieldy when getting bigger, for the sheer number of states they produce.
Hierarchical state machines as defined by the statechart notation alleviate this problem by giving an additional structural element through grouping states into hierarchies.
The "State" pattern in its original form is not capable of modeling state hierarchies. The Macho class library extends the concept with this possibility, keeping the properties of simplicity (there possible) and tool independence from its inspiration.
The class library as such does not need to be installed. Just include the header file Macho.hpp to make use of it. Prerequisite however is a C++ compiler with sane support for templates.
Included are the example state machines HelloWorld, Example, Microwave and Test. To make the examples run just compile them in the directory they are in, for example:
# GCC
g++ -o microwave Microwave.cpp # MSVC7
cl /EHsc Microwave.cpp
The following descriptions assume at least passing knowledge of the statechart notation. For more information see Wikipedia.
Starting point is the GoF "State" design pattern. The "State" pattern models states with classes. State transitions are done by instantiating objects of those classes. The class constructors and destructors take on the role of entry and exit actions. Class attributes represent state variables. Events are dispatched by calling methods on state objects.
For states to compose a hierarchy, substates must be able to take over the guards, actions and transitions of superstates, selectively redefining them where necessary. There exists a mechanism in C++ allowing exactly this on method level: polymorphism through class inheritance.
Modeling the substate-superstate relation with class inheritance is problematic however: the use of constructors and destructors as entry and exit actions is not possible anymore, neither is keeping state variables in objects.
The reason is that superclasses are constituent parts of their subclasses, meaning object construction or destruction will trigger all involved class constructors or destructors and initialize or destroy all data members.
This runs counter to the semantics of entry/exit actions and state variables, where a transition between sibling substates should not trigger superstate entry/exit actions nor destroy superstate state variables.
Our solution to this problem is to use explicit methods for state entry and exit, being called in the correct sequence on state transitions. State variables are kept in separate state specific data structures which have a lifecycle consistent with the hierarchy of states.
Before diving into implementation details we first define some terminology to be used from here on. By state we mean the class defined to represent a particular state. A state machine is a set of states having a common top state class. A state machine instance maintains a current state to which events can be dispatched. A state machine's event protocol is the set of events that are understood by the machine.
Furthermore we will use extracts from the simple state machine defined in the
file Example.cpp for illustration. It is a good idea
to have a look at this file now if you are reading this for the first time. It
contains a state machine with four states: a top state, a
superstate Super, two substates StateA
and StateB, and a small test run. Each state has entry/exit
actions, the top state and StateA have state variables and the
event protocol consist of two events, event1
and event2.
The top state of the example is defined by using the
macro TOPSTATE:
TOPSTATE(Top) {
...
};
This sample defines the top state class Top.
Every state machine must have a top state. The top state's interface defines the machine's event protocol: only the public virtual methods of the top state can be event handlers.
All other states of a machine are direct or indirect substates of this top
state. The top state has a typedef alias TOP known to
all states in the machine.
Substates are defined by the SUBSTATE macro:
SUBSTATE(Super, Top) {
...
};
This example defines the class Super as substate of
class Top. The macro parameters are the new substate's name and the
name of its superstate. The superstate can be the top state or any other substate.
Why are we using macros in Macho? The reason being only convenience, you'll most
likely agree with their use if you look at the definitions of the macros
TOPSTATE and SUBSTATE:
#define TOPSTATE(TOP) \
struct TOP : public Macho::Link< TOP, Macho::Root< TOP > >
#define SUBSTATE(STATE, SUPERSTATE) \
struct STATE : public Macho::Link< STATE, SUPERSTATE >
The code is more readable and actually less error-prone (notice the multiple
use of macro parameters in TOPSTATE and SUBSTATE)
by using macro expansion.
Another macro is STATE(S) which MUST appear in every state
body:
SUBSTATE(Super, Top) {
STATE(Super)
...
};
The macro parameter is the state's name again. With this macro some definitions are provided in the class body (a constructor for instance). For more details look up the macro's definition.
Every state must be instantiable (the top state as well)! This means states may not have pure virtual methods!
A state's (optional) entry and exit actions are defined this way:
TOPSTATE(Top) { ... private: void entry(); void exit(); void init(); }; void Top::entry() { ... } void Top::exit() { ... } void Top::init() { ... }
The methods entry and exit of a state are called
upon transitioning into or out of it. The sequence of calls for a state
transition is as follows:
First the exit action of the current state is called and then those of its superstates (in bottom up order), up to and excluding the first superstate that is also superstate of the new state.
Then entry actions of the superstates of the new state are called, top down from the first superstate that is not also a superstate of the previous state. Finally the entry action of the new state is called.
From this follows that self transitions (where the new state is the same state as the current state) will trigger that state's exit and entry actions.
Entry/exits actions may NOT initiate new state transitions! Doing otherwise will cause an assertion.
The method init defines a special kind of entry action: Upon
state transition, entry actions of the new state and its superstates are
triggered; init however is called only on the one state the
transition actually goes to, after its entry action is performed (please see the
section History for an exception if the state has history
enabled).
With init the initial transitions of states are implemented,
where a superstate immediately enters a substate on direct entry (like top state
in Example). This implies that init actions ARE in fact allowed to
initiate state transitions (but only to substates). Please note that
transitioning to a state in init will trigger that new
state's init method in turn.
It is recommended that you put the methods exit,
entry and init into the private section of your
classes, at least for the top state, so that they may not be called from third
parties.
Constructors or destructors can not be used for state classes. If you need to make initializations, do them in your entry actions or box constructors!
State variables are data members that have the lifetime of their associated states. For superstate variables this means that they exist as long as the machine is in that state or in any of its substates. Substates have full access to state variables of their superstates.
State variables allow states to accumulate information from the events they have received in the past. They can also be used to parametrize a new state to be entered by initializing them accordingly.
Since every state is a substate of the top state, top state variables are accessible to all states of a machine.
Here is an example of a definition of state variables in Macho:
TOPSTATE(Top) { struct Box { Box() : myData(0) {} Box(long l) : myData(l) {} long myData; }; STATE(Top) ... };
State variables are contained in a data type named Box nested
into the state class (the type name Box is mandatory). The box
definition must appear BEFORE the use of the STATE macro.
The box must be default constructable (has a default constructor), but may have other constructors, too. Apart from this the box can take on any form of a regular C++ type. It could even be just a simple typedef:
TOPSTATE(Top) { typedef int Box; STATE(Top) ... };
A box is created before its state is entered (before the call
to entry) and destroyed after the state is left (after the call
to exit).
A state's box is accessed by calling the method box, which returns a
reference to the state's box object:
void StateA::event1(int i) { ... cout << box().myData; ... }
Superstate boxes are accessed by qualifying the box method with
the superstate's name:
void StateB::event2(long l) { ... cout << TOP::box().myData; ... }
Here the top state's box is requested.
As mentioned the event protocol of a state machine is defined by its top state interface:
TOPSTATE(Top) {
...
virtual void event1(int i) {}
virtual void event2(long l) {}
...
};
The Example state machine understands the events event1 and
event2 (the event name is the same as the name of its event handler
method). Events may have arbitrary parameters.
Event handlers are implemented as ordinary C++ methods:
SUBSTATE(StateA, Super) { ... void event1(int i); ... }; void StateA::event1(int i) { ... }
In this fragment the event handler for event1 in
state StateA is defined. Inside the method the guards, actions and
state transitions are implemented.
The top state event handlers define the default behaviour for the whole state machine. If there is no meaningful implementation for an event handler at top level, the handler could either
assert(false).State transitions are made by calling the method setState inside
an event handler:
void StateA::event1(int i) { ... setState<StateB>();
}
The template parameter to setState is the new state to be
entered.
The state transition takes place AFTER control flow leaves the event
handler. This means it is possible to do meaningful work even after
calling setState (all involved objects are still alive). It is
however not allowed to call setState multiple times with different
states in a single event handler run (this will assert on you)!
There is an optional parameter to setState:
void Top::init() { setState<StateA>(new StateA::Box(44)); }
By creating a state's box prior transition it is possible to parametrize a new state by initializing its box suitably. The given box will then be used for the new state.
We recommend that you do not define event handlers inline to your state class
definitions. The reason is that C++ needs a complete definition of a class to be
used as template parameter, and that may not be the case for the classes you
want to setState to in inline event handlers.
It is possible for a superstate to remember a previously entered substate. The next (direct) entry into the superstate will then reenter the remembered substate.
History is activated for a state by use of the macro HISTORY or
DEEPHISTORY:
SUBSTATE(Super, Top) { STATE(Super) HISTORY() ... };
Use of HISTORY selects the shallow history strategy, whereby
only direct substates of a superstate are remembered. DEEPHISTORY
remembers even the substates of substates.
History is only evaluated for a state if it is the direct target of a state transition. If the transition goes to one of its substates, that substate is entered without checking superstate history.
If a state has history enabled, its init method will not be
called if a history state is available. This is because history takes precedence
over initial transitions. This means that for a state with history
init is called only the very first time the
state is entered.
It is possible to ignore history information for a state transition however
by using the method setStateDirect instead
of setState:
void StateA::event1(int i) { ... setStateDirect<Super>();
}
This example will enter Super without checking its history,
calling init instead.
Keep in mind that state boxes are not remembered but rather instantiated anew for reentered history states!
A state machine instance is created by instantiating a machine object:
Macho::Machine<Example::Top> m();
The class Machine is a template class in namespace
Macho; its template parameter is the top state of the state
machine to be run.
The top state is immediately entered and initialized upon machine creation
(entry and init are called on it).
Upon destruction of the machine object the current sub- and superstates up to and including the top state are exited correctly and their boxes destroyed.
The Machine constructor takes as optional parameter a box for
the top state:
Macho::Machine<Example::Top> m(new Example::Top::Box(11));
This allows parametrization of a state machine on startup.
It is also possible to read the top state box of a running machine by calling
the box method of Machine:
cout << m.box().myData;
The method box of class Machine returns
a const reference to the top state box. It is therefore not
possible to change box data (you should really do this from INSIDE your machine)!
A state's history for a particular machine instance can be cleared by calling
the state's static method clearHistory with the machine object as
parameter:
Example::Super::clearHistory(m);
This statement resets history information for Super inside
machine m (without affecting history of Super's
substates however).
Another static method of state classes allows testing for the current state of a machine object:
assert(Example::StateA::isCurrent(m));
The method isCurrent returns true if the given
machine object is in the selected state or any of its substates at that
moment (use isCurrentDirect to ignore substates).
Events are dispatched to a machine by calling event handlers through the machine object's arrow operator:
m->event1(42);
m->event2(43);
The machine will call the corresponding event handler of its current state.
The overloading of the arrow operator is the reason why the whole event protocol must be declared in the top state class: only then can all events be dispatched by the machine object.
For more detailed information about Macho please consult the implementation source code and included samples.
| Q: | Why C++? |
| A: | Macho needs good template support. |
| Q: | I'm getting weird compiler errors... |
| A: | Check
if you have
|
| Q: | How fast is this stuff? |
| A: | Benchmarking the nontrivial state machine Test (without printing) gives me around 1.000.000 state transitions per second on my 2.4 GHz P4 machine. |
| Q: | Does it work with the Visual Studio 6 compiler? |
| A: | At present Macho can not be compiled in the MSVC6 C++ dialect. I'm looking into it, but this has LOW priority! |