Re: [boost] [msm]exit pseudo state and event

Subject: Re: [boost] [msm]exit pseudo state and event
From: Christophe Henry (christophe.j.henry_at_[hidden])
Date: 2011-07-19 18:30:03

>>>> > Off the top of my head, how about the functor Else_ below?
>>>> > Else_ <Guard1, Guard2, Guard3, ... >
>>>>
>>>> For else, we can use the fact that msm tries guards from the bottom of
>>>> the
>>>> table to the top:
>>>>
>>>> // else clause
>>>> Row< state1, ev, state2, none, none>,
>>>> // if clause
>>>> Row< state1, ev, state2, none, Or_<...> >
>>>> // more if clauses
>>>> ...
>>>
>>> Fantastic!
>>> This has 2 advantages and 1 disadvantage.
>>>
>>> Advantages:
>>> One is simple enough. The other is good performance.
>>> It doesn't need evaluating the guard functor twice.
>>> I was about to consider how to cache the functor's result.
>>>
>>> Disadvantage:
>>> As you might expect, the evaluating direction is opposite of my 
>>> intuition.
>>>
>>> Are there any reasons of this implementation?
>>
>> There are.
>> It's an implementation detail which dates back to the early phases of the
>> library (start + 3 months).
>> The reason is that msm adds, for performance, rows for submachines at the
>> end of the table (one for every event found in the submachine), then 
>> starts
>> processing from the bottom to give these rows a higher priority (UML 
>> rule).
>
> I agree.
> But this UML rule should only apply the submachine structure.
> I describe more details below.

Sure but msm implements all those cases.

>> At the same time, the library also uses the table to generate state ids
>> (http://svn.boost.org/svn/boost/trunk/libs/msm/doc/HTML/ch06s03.html).
>> I could do the opposite (extra rows at the beginning) but it'd be pretty
>> hard to guess after this what the state id you get in no_transition means
>> (it's already hard enough to explain the current order, imagine if I had 
>> to
>> explain you have to count in reverse order...).
>
> I agree.
> State id ordering shouldn't be changed.
>
>> And as the else feature is not in the Standard, it seemed ok.
>> In these early phases I didn't have, like now, different steps of 
>> transition
>> table processing.
>> This old decision could deserve some revisiting, but:
>> - there are a few corner cases which could make a change pretty tricky
>> (pseudo entry/exit states)
>
> I think it isn't corner case.
> Consider else_usecase.png.(attached file)
> When we implement choice pseudo state, we often use the else guard.
> And when we implement this statemachine using msm, we use extra event like 
> the
> substatemachine case.
> See else_usecase_converted.png.
>
> IMHO, we often meet with this case.

I'm afraid you misunderstood me. I didn't write that a choice pseudo-state 
is a corner case but that the change you request has a few corner cases to 
consider if the implementation is to work.

>> - some people already use the current implementation state, so it'd break
>> their code, I'd need some compile-time configuration, which would get 
>> hidden
>> deep in the already big doc. The default would still be the current 
>> order.
>
> Current documents refer to only state id order.
> http://www.boost.org/doc/libs/1_47_0/libs/msm/doc/HTML/ch06s03.html
>
> They don't refer to reverse transition evaluation. I believe it's
> still implement detail.
> This meas we have a chance to change the evaluation order.

No, this is documented 
(http://www.boost.org/doc/libs/1_47_0/libs/msm/doc/HTML/ch06.html): "the 
guard conditions are tried in reverse order of their transition definition 
in the transition table" and I know of a few people using this.

> Here is my solution.
>
> See forward_eval.patch.
>
> In dispatch_table.hpp,
> If the state is submachine state then front_row and Fsm::frow<State,
> Event> are same.
> And it appears only front element.
> The evaluating direction changes by this information.
> After applying this patch, 14_GuardEvalOrder.cpp(attached file)
> outputs the result below.
>
> Guard1_1
> Guard1_2
> Guard1_3
> Guard1_4
> Guard1
> Guard2
> Guard3
> Guard4
>
> Both evaluation priority and order are OK :)
>
> But there is another problem that relate to internal transitions.
> http://www.boost.org/doc/libs/1_47_0/libs/msm/doc/HTML/ch03s02.html#d0e807
>
> The document said internal transitions have the highest priority(as
> mandated by the UML standard).
> I checked UML standard. But I couldn't find this rule.
> (OMG Unified Modeling LanguageTM (OMG UML) Superstructure Version 2.3)
>
> I'm guessing you mean below.
>
> 15.3.12 StateMachine (from BehaviorStateMachines)
>
> === begin quotation ===
> Conflicting transitions
> It was already noted that it is possible for more than one transition
> to be enabled within a state machine. If that happens,
> then such transitions may be in conflict with each other. For example,
> consider the case of two transitions originating
> from the same state, triggered by the same event, but with different
> guards. If that event occurs and both guard conditions
> are true, then only one transition will fire. In other words, in case
> of conflicting transitions, only one of them will fire in
> a single run-to-completion step.
> Two transitions are said to conflict if they both exit the same state,
> or, more precisely, that the intersection of the set of
> states they exit is non-empty. Only transitions that occur in mutually
> orthogonal regions may be fired simultaneously. This
> constraint guarantees that the new active state configuration
> resulting from executing the set of transitions is well-formed.
> An internal transition in a state conflicts only with transitions that
> cause an exit from that state.
>
> Firing priorities
> In situations where there are conflicting transitions, the selection
> of which transitions will fire is based in part on an
> implicit priority. These priorities resolve some transition conflicts,
> but not all of them. The priorities of conflicting
> transitions are based on their relative position in the state
> hierarchy. By definition, a transition originating from a substate
> has higher priority than a conflicting transition originating from any
> of its containing states.
> The priority of a transition is defined based on its source state. The
> priority of joined transitions is based on the priority
> of the transition with the most transitively nested source state.
> In general, if t1 is a transition whose source state is s1, and t2 has
> source s2, then:
> * If s1 is a direct or transitively nested substate of s2, then t1
> has higher priority than t2.
> * If s1 and s2 are not in the same state configuration, then there is
> no priority difference between t1 and t2.
> Transition selection algorithm
> The set of transitions that will fire is a maximal set of transitions
> that satisfies the following conditions:
> * All transitions in the set are enabled.
> * There are no conflicting transitions within the set.
> * There is no transition outside the set that has higher priority
> than a transition in the set (that is, enabled transitions
> with highest priorities are in the set while conflicting transitions
> with lower priorities are left out).
> This can be easily implemented by a greedy selection algorithm, with a
> straightforward traversal of the active state
> configuration. States in the active state configuration are traversed
> starting with the innermost nested simple states and
> working outwards. For each state at a given level, all originating
> transitions are evaluated to determine if they are
> enabled. This traversal guarantees that the priority principle is not
> violated. The only non-trivial issue is resolving
> transition conflicts across orthogonal states on all levels. This is
> resolved by terminating the search in each orthogonal
> state once a transition inside any one of its components is fired.
> === end quotation ===

Yes, it is probably this.

> See smd1.png , smd2.png and smd3.png. These are my understanding.
> And I wrote internal_continue.patch.
> This modification is to continue evaluating process if evaluated row
> is internal transition.
>
> If my understanding is right, we don't need the following expression.
>
> struct Empty : public msm::front::state<>
> {
>    struct internal_transition_table : mpl::vector<
>           a_internal < cd_detected , Empty, &Empty::internal_action >
>    > {};
> };

Are you saying we don't need internal transitions? I'd disagree then. 
They're the best way to implement taking action without changing state.

> You can check this behavior using 15_MultiInternalTransition.cpp.
>
>>> Modifying to the opposite direction is easy?
>>
>> Unfortunately not, so considering the gain and the time it'd cost, it's
>> unlikely to change until I have nothing else to implement in the library,
>> which could take some time ;-)
>
> I also did the tests.
> To pass the all tests, it requires some changes. See msm_test.patch.
> The changes are only ordering of rows.
> Internal transitions move up in the same start state group.
> These changes doesn't effect the state ids.
>
> After 3 patches are applied, all tests passed.

Well, yes, if you change the test, of course it will work ;-)

> I tested on the following environments.
> g++ (Ubuntu 4.4.3-4ubuntu5) 4.4.3
> Microsoft Visual C++ 2008 Express Edition SP1

Unfortunately my tests are not sufficient, it's something I want to take 
care of soon.

> Could you check my patches?

I did. I'm not sure I fully get what you're trying but I don't think it will 
work.
Again, there are quite some cases to think about.
Let's take one, to show why the priority given to inner is important. See 
conflict_and_exit.
If I believe your example in smd1, there is no priority. Then, the exit 
points cannot work any more (yes, transitions leading to exit points are 
inner transitions).

Frankly, I don't see the point in doing such change:
- it will break people's code
- it will break a documented implementation
- I doubt it will work in all cases
- it is likely to be slower (more if statements)
- it will not be Standard-conform. To do this, I need a really good reason

And this for no bugfix and no added feature, simply to be able to write:

// else clause
Row< state1, ev, state2, none, none>
// if clause
Row< state1, ev, state2, none, Or_<...> >

instead of:

// if clause
Row< state1, ev, state2, none, Or_<...> >
// else clause
Row< state1, ev, state2, none, none>

No, this does not look appealing to me.
If you're at implementing stuff for msm, I might consider another solution:
- a typedef in the front-end, which will change the reading order of the 
table, no to break other's code
- replacing in dispatch_table reverse_fold by fold to change the reading 
order if typedef is requesting it.
- adding rows (frows) at the top instead of bottom.

This leaves us with the hard part, reading the ids in the correct order. For 
this, we need to find out which one of the transition table transformations 
is the correct one. I didn't look at it yet, but I imagine it could be 
tricky. Then use it to read the id from up to down.

I understand you like this feature, but I need to also consider the wishes 
of those who requested me very useful features some time ago and are waiting 
for them.
So it's also my job to allocate my scarce development time (my evenings) 
where I think it will be the most useful.
For example, the junction pseudo-state you requested seems to me to have a 
much higher gain for all of us.

Regards,
Christophe