September 2010

A pattern for GUI programming

This document summarizes the way LightWidgets are intended to be used. The style for GUI programming is based on PluggableViews in MVC and PluggableMorphs in Morphic. The main idea is to have a reusable set of standard widgets that can be customized when used by application developers. There is a strict separation between views and models. Models don't know about views, they are never aware of them. Views know about their model and update it directly. Therefore, there's no need for views to trigger events, and they don't do it.

This description is not only conceptual, or theoretic. The rules described here are to actually be followed.

GUIs are built by composing widgets. The main view is a subclass of CompositeLW. There is a complete separation between Model and Views.

Rule 1. Models should never include GUI code

They must be completely unaware of possible Views that operate on them. There could be at any time any number of different views active on the same model. They could belong to different technologies or frameworks. They could even be remote and run on a different computer. There could be no view at all. For example, the model could be driven by scripts or reside on a server and receive external commands. However, this document will only describe local LightWidgets GUIs.

Rule 2. Views should never include any model code

The view could be replaced anytime with a different one. Besides, a model should be able to run without any GUI at all. So any logic that belongs in the model but is included in the GUI will eventually be missing. The Views should query and modify Models only through public protocols, called 'Inquiries' and 'User Commands'.

Rule 3. Views know about the model they operate on

Views have an instance variable to hold their model. They can query Model Inquiries when needed. They can also issue User Commands when appropriate. Models are usually subclasses of ActiveModel. Let's consider a small example. We are building a GUI to operate on some Person objects. We'll consider an EntryField for the birthday of aPerson. LightWidget includes the following instance variables:

target : Holds the Model. For our example, the model of the EntryField would be the Person. We call it target, because sometimes it might be a view
aspect : It is a symbol, the getter for the aspect we are showing. In this case it would be #birthday.
aspectAdaptor : It is a symbol, a message that is sent to the aspect to adapt it to the widget. As the widget is an EntryField and the aspect is a Date, the aspectAdaptor could be #asString. This relies the Model from the need to provide an appropriate getter for each kind of possible widget for each attribute.
action : The action is the setter used to update the aspect on the model. In this example, it is #birthday:.
actionAdaptor : It is used to adapt the value the user entered in the widget for use as an argument of action. In this example it could be #asDate.

It is usually a good idea to initialize model aspects with reasonable defaults, and avoid nil values. This saves a lot of #ifNil: messages in the gui.

Rule 4. View Structure

A Model could have a tree-like structure. It could be composed of other Models. This is not mandatory.

Views always have a tree-like structure. The leaves are simple widgets. The internal nodes are CompositeLWs. They can all share the same model, or they could could use different parts of the bigger model. Anyway, they are always customized with the aspec and action.

Rule 5. GUI construction

The construction of the Views tree and the customization of each widget is done by a main view. The main view also specifies how the Views are notified of model changes for updating.

Rule 6. Instance variables in views

Additional instance variables in GUIs are of two kinds: They can be uses to hold sub-views, or to hold 'Model Extensions'. Possible uses of Model Extensions include:

Holding information that can be obtained from the aspect, but that could be expensive and it makes sense to cache. For example, our EntryField could hold an array if indices of word starts and ends or some other internal detail.
Holding state that is meaningful for the widget, but that it doesn't make any sense to keep in the model. An example could be the cursor position in our EntryField. Others could be visual options, such as a graph type or graph style for an application generated graph.
Not-yet-commited information, entered by the user, but awaiting for OK / Cancel.

In general, Model Extensions usually are re-fetched from the model, or re-set to default values when the model changes.

Rule 7. View updating because of model changes

Any widget (in fact, any Morph) can redraw itself when needed, with the #changed method. But when there is a change in the Model, the views must be updated appropriately. All the Model Extensions must be updated, and all sub-views must be updated too.

When there is a change in the Model, the Views must receive the #modelChanged message. A main view (i.e. a view that is not subview of another view with the same Model) must send itself #beMainViewOn: on construction. This does 'target when: #selfChanged send: #safeModelChanged to: self'. The Model must trigger event #selfChanged when appropriate. #safeModelChanged will eventually update all subviews recursively. So only a main view should receive the #selfChanged event. Models are usually subclasses of ActiveModel, to use the more advanced events implementation there.

This is the implementation of #beMainViewOn: . This message should be used to set the model of a main view.

beMainViewOn: aModel
   "We are a main view on aModel.
   This means:
       - aModel is a real model, i.e. not a widget.
       - no aspect or aspectAdaptor. We show the whole thing.
       - no action or actionAdaptor. There is no main action.
       - we must update ourselves on #selfChanged event"

   self target: aModel aspect: nil aspectAdaptor: nil modelChangeEvent: #selfChanged

The main update method is #modelChanged. #safeModelChanged is only to guarantee that the update is done in the User Interface process, in the inter-cycle paus. The implementation of #modelChanged at LightWidget is:

modelChanged
   "The model changed is some way.
   This is usually the pace to call #targetAspect to fetch the current value of the aspect from the
       model, and to store it in some Model Extension.
   We must update all Model Extension instance variables with values from the model (i.e. target)
       or with appropriate defaults.
   We must update ourselves and all subviews to reflect the model's new state"

   self updateView

#modelChanged must be reimplemented in classes with model extensions. Check the implementors to see how they work.

After updating the model and model extensions, #updateView is called. This is the implementation at LightWidget:

updateView
   "The model or some Model Extension changed is some way.
   We must update ourselves to reflect the new state.

   This is the place to update secondary Model Extensions or any other state that must be updated
   after model or Model Extension change.

   This method is usually reimplemented in CompositeLWs, to update subviews.

   The subviews should be sent one of the following messages:
       target:
       target:aspect:
       target:aspect:aspectAdaptor:
       target:aspect:aspectAdaptor:aspectChangeEvent:
   to update their model and do a full update, as triggered by #modelChanged"

   self changed

Warning: Never implement other methods like #updateViews. If for performance reasons the updating of subviews must be splitted in parts, then the views and subviews must be restructured accordingly. Then, each part can be updated as a whole with the #updateView method. Each part can be updated by more specific model change events, or alternatively, they might be set different submodels. Both options are described below.

The update of widgets should never trigger the action of the widget.

Rule 8. View updating because of Model Extension changes

If the target of a widget is another widget, the action is a User Command on the target widget. These methods should not update the model, because if this was the case, the target should be the model and not the widget. Therefore, User Command methods in widgets can only update Model Extensions or trigger view actions, such as opening new views, etc. If they update Model Extensions they should call #updateView, so the change is shown in the widget and its subviews.

sampleUserCommand: data
   modelExtension1 := data.
   self updateView
   "Must call updateView because the model didn't change, and it will not trigger any change event"

Rule 9. Subview updating because of submodel changes

If the model has a tree-like structure, its view will send #beMainViewOn: aSubModel to some subviews with a part of the model as the argument. In this case, subviews will need to be notified of the events of their own models. This is because the submodel might trigger the #selfChanged event, and only the views on it should be updated. Views on the bigger model don't need to be updated. This is good for performance when having complex models and views.

Rule 10. Subview updating because of model minor changes

There is another reason for subviews receiving event notifications. A model could trigger a more specific #someAspectChanged event and NOT the main #selfChanged event. This could be done to avoid superfluous and extensive views updating. In this case, some specific view on the view tree should receive the #updateView message, and only the widgets that are part of it will be updated.

So, the owning view should send #target:aspect:aspectAdaptor:modelChangeEvent: to these subviews. The implementation is:

target: aModel aspect: aSymbol aspectAdaptor: anotherSymbol modelChangeEvent: eventSymbol
   "Widgets are notified of model changes by being sent #modelChanged.
   This happens when:
       - The widget is given a new model (or target widget), aspect or aspect adaptor
       - An owner view is updated

   In addition, main views are updated from model events. See #beMainViewOn:

   But other widgets might update on more specific events from the model. This is useful to
   update only a small subview, and not the whole main view.

   This message is sent to such widgets, to set this specific event.

   Warning:
   When models change, they should trigger just one event.
   It might be #selfChanged (the most general one) or a more specific one.
   But it should not trigger more than one event for each change."

   self target: aModel aspect: aSymbol aspectAdaptor: anotherSymbol.
   target when: eventSymbol send: #safeModelChanged to: self

Warning: When a model triggers more specific change events we must make sure some widget will be notified of them. Otherwise, those changes could not be shown to the user.

Rule 11. Accessing views

Nobody should ever query a widget for value or status. A widget should not even query itself for current value or such. The last value or state entered by the user should be stored in the model and/or Model Extensions. When needed, it should be retrieved from there. The only legitimate accesses to subviews are in #initialize and in #updateView. Check implementors of #updateView.

Rule 12. Model updating

Views DO NOT trigger events. This is not "Event Oriented Programming". This is Object Oriented Programming. The model is updated using the action and the optative actionAdaptor. Methods that react to user activity should update the model by just using the action, a simple message. They are not allowed to ask the model for some other object to work on it. They are not allowed to send other messages to the model. They are allowed to modify Model Extensions. If they do, they should also send 'self modelChanged' because an action might not modify the model and therefore there could not be a model change event. See ButtonLW>>mouseUp: for an example of this.

If you ever feel the need to update the object answered by the aspec, instead of sending a new value to the model (ivar target), it is because that aspect should be the real model.

Rule 13. GUI building

Main views know about their subviews. Therefore it's them, in theire #initialize method, who build the subviews and customizes them. Views are created before assigning target or model to the main view. Afterwards, the model or target is set, and #modelChanged is called. As seen before, this will set the model or target of all subviews recursively.

Misc. notes

I believe nobody should do #modelChanged, but only #safeModelChanged. Think a bit about this. Maybe if we're certain we're in the UI process, #modelChanged is ok...

If a visual detail like #fontColor: in a LabelLW is updated, after updating the ivar, the widget should do 'self changed'. Check the code to see that it is actually done!

Thanks for reading.