| snitfaq(n) | Snit's Not Incr Tcl, OO system | snitfaq(n) |
snitfaq - Snit Frequently Asked Questions
This is an atypical FAQ list, in that few of the questions are frequently asked. Rather, these are the questions I think a newcomer to Snit should be asking. This file is not a complete reference to Snit, however; that information is in the snit man page.
Snit is a framework for defining abstract data types and megawidgets in pure Tcl. The name "Snit" stands for "Snit's Not Incr Tcl", signifying that Snit takes a different approach to defining objects than does Incr Tcl, the best known object framework for Tcl. Had I realized that Snit would become at all popular, I'd probably have chosen something else.
The primary purpose of Snit is to be object glue--to help you compose diverse objects from diverse sources into types and megawidgets with clean, convenient interfaces so that you can more easily build your application.
Snit isn't about theoretical purity or minimalist design; it's about being able to do powerful things easily and consistently without having to think about them--so that you can concentrate on building your application.
Snit isn't about implementing thousands of nearly identical carefully-specified lightweight thingamajigs--not as individual Snit objects. Traditional Tcl methods will be much faster, and not much more complicated. But Snit is about implementing a clean interface to manage a collection of thousands of nearly identical carefully-specified lightweight thingamajigs (e.g., think of the text widget and text tags, or the canvas widget and canvas objects). Snit lets you hide the details of just how those thingamajigs are stored--so that you can ignore it, and concentrate on building your application.
Snit isn't a way of life, a silver bullet, or the Fountain of Youth. It's just a way of managing complexity--and of managing some of the complexity of managing complexity--so that you can concentrate on building your application.
Snit 1.3 requires Tcl 8.3 or later; Snit 2.2 requires Tcl 8.5 or later. See SNIT VERSIONS for the differences between Snit 1.3 and Snit 2.2.
Snit is part of Tcllib, the standard Tcl library, so you might already have it. It's also available at the Snit Home Page, http://www.wjduquette.com/snit.
Snit is unique among Tcl object systems in that it is based not on inheritance but on delegation. Object systems based on inheritance only allow you to inherit from classes defined using the same system, and that's a shame. In Tcl, an object is anything that acts like an object; it shouldn't matter how the object was implemented. I designed Snit to help me build applications out of the materials at hand; thus, Snit is designed to be able to incorporate and build on any object, whether it's a hand-coded object, a Tk widget, an Incr Tcl object, a BWidget or almost anything else.
Note that you can achieve the effect of inheritance using COMPONENTS and DELEGATION--and you can inherit from anything that looks like a Tcl object.
Using Snit, a programmer can:
The current Snit distribution includes two versions, Snit 1.3 and Snit 2.2. The reason that both are included is that Snit 2.2 takes advantage of a number of new features of Tcl 8.5 to improve run-time efficiency; as a side-effect, the ugliness of Snit's error messages and stack traces has been reduced considerably. The cost of using Snit 2.2, of course, is that you must target Tcl 8.5.
Snit 1.3, on the other hand, lacks Snit 2.2's optimizations, but requires only Tcl 8.3 and later.
In short, if you're targetting Tcl 8.3 or 8.4 you should use Snit 1.3. If you can afford to target Tcl 8.5, you should definitely use Snit 2.2. If you will be targetting both, you can use Snit 1.3 exclusively, or (if your code is unaffected by the minor incompatibilities between the two versions) you can use Snit 1.3 for Tcl 8.4 and Snit 2.2 for Tcl 8.5.
To always use Snit 1.3 (or a later version of Snit 1.x), invoke Snit as follows:
package require snit 1.3
To always use Snit 2.2 (or a later version of Snit 2.x), say this instead:
package require snit 2.2
Note that if you request Snit 2.2 explicitly, your application will halt with Tcl 8.4, since Snit 2.2 is unavailable for Tcl 8.4.
If you wish your application to always use the latest available version of Snit, don't specify a version number:
package require snitTcl will find and load the latest version that's available relative to the version of Tcl being used. In this case, be careful to avoid using any incompatible features.
To the extent possible, Snit 2.2 is intended to be a drop-in replacement for Snit 1.3. Unfortunately, some incompatibilities were inevitable because Snit 2.2 uses Tcl 8.5's new namespace ensemble mechanism to implement subcommand dispatch. This approach is much faster than the mechanism used in Snit 1.3, and also results in much better error messages; however, it also places new constraints on the implementation.
There are four specific incompatibilities between Snit 1.3 and Snit 2.2.
dog spot ;# Explicit naming
set obj1 [dog %AUTO%] ;# Automatic naming
set obj2 [dog] ;# Implicit naming
In Snit 2.2, type commands are defined using the namespace ensemble
mechanism; and namespace ensemble doesn't allow an ensemble command
to be called without a subcommand. In short, using namespace
ensemble there's no way to support implicit naming.
All is not lost, however. If the type has no type methods, then the type command is a simple command rather than an ensemble, and namespace ensemble is not used. In this case, implicit naming is still possible.
In short, you can have implicit naming if you're willing to do without type methods (including the standard type methods, like $type info). To do so, use the -hastypemethods pragma:
pragma -hastypemethods 0
A hierarchical method is an instance method which has subcommands; these subcommands are themselves methods. The Tk text widget's tag command and its subcommands are examples of hierarchical methods. You can implement such subcommands in Snit simply by including multiple words in the method names:
method {tag configure} {tag args} { ... }
method {tag cget} {tag option} {...}
Here we've implicitly defined a tag method which has two subcommands,
configure and cget.
In Snit 1.3, hierarchical methods could be called in two ways:
$obj tag cget -myoption ;# The good way
$obj {tag cget} -myoption ;# The weird way
In the second call, we see that a hierarchical method or type method is
simply one whose name contains multiple words.
In Snit 2.2 this is no longer the case, and the "weird" way of calling hierarchical methods and type methods no longer works.
In Snit 2.2, hierarchical methods and type methods are no longer simply ones whose name contains multiple words; in the above example, the list returned by $obj info methods would include tag but not tag configure or tag cget.
This is extremely convenient. However, it also means that commands defined in the parent namespace, ::mypackage:: can block the type's access to identically named commands in the global namespace. This can lead to bugs. For example, Tcllib includes a type called ::tie::std::file. This type's code calls the standard file command. When run with Snit 2.2, the code broke-- the type's command, ::tie::std::file, is itself a command in the type's parent namespace, and so instead of calling the standard file command, the type found itself calling itself.
Yes.
This is extremely useful when a type is defined as part of a larger package, and shares a parent namespace with the rest of the package; it means that the type can call other commands defined by the package without any extra work.
This feature depends on the new Tcl 8.5 namespace path command, which is why it hasn't been implemented for V1.x. V1.x code can achieve something similar by placing
namespace import [namespace parent]::*in a type constructor. This is less useful, however, as it picks up only those commands which have already been exported by the parent namespace at the time the type is defined.
A full description of object-oriented programming is beyond the scope of this FAQ, obviously. In simple terms, an object is an instance of an abstract data type--a coherent bundle of code and data. There are many ways to represent objects in Tcl/Tk; the best known examples are the Tk widgets.
A Tk widget is an object; it is represented by a Tcl command. The object's methods are subcommands of the Tcl command. The object's properties are options accessed using the configure and cget methods. Snit uses the same conventions as Tk widgets do.
In computer science terms, an abstract data type is a complex data structure along with a set of operations--a stack, a queue, a binary tree, etc--that is to say, in modern terms, an object. In systems that include some form of inheritance the word class is usually used instead of abstract data type, but as Snit doesn't implement inheritance as it's ordinarily understood the older term seems more appropriate. Sometimes this is called object-based programming as opposed to object-oriented programming. Note that you can easily create the effect of inheritance using COMPONENTS and DELEGATION.
In Snit, as in Tk, a type is a command that creates instances -- objects -- which belong to the type. Most types define some number of options which can be set at creation time, and usually can be changed later.
Further, an instance is also a Tcl command--a command that gives access to the operations which are defined for that abstract data type. Conventionally, the operations are defined as subcommands of the instance command. For example, to insert text into a Tk text widget, you use the text widget's insert subcommand:
# Create a text widget and insert some text in it.
text .mytext -width 80 -height 24
.mytext insert end "Howdy!"
In this example, text is the type command and .mytext is the instance command.
In Snit, object subcommands are generally called INSTANCE METHODS.
Snit allows you to define three kinds of abstract data type:
A snit::type is a non-GUI abstract data type, e.g., a stack or a queue. snit::types are defined using the snit::type command. For example, if you were designing a kennel management system for a dog breeder, you'd need a dog type.
% snit::type dog {
# ...
}
::dog
%
This definition defines a new command (::dog, in this case) that can be used to define dog objects.
An instance of a snit::type can have INSTANCE METHODS, INSTANCE VARIABLES, OPTIONS, and COMPONENTS. The type itself can have TYPE METHODS, TYPE VARIABLES, TYPE COMPONENTS, and PROCS.
A snit::widget is a Tk megawidget built using Snit; it is very similar to a snit::type. See WIDGETS.
A snit::widgetadaptor uses Snit to wrap an existing widget type (e.g., a Tk label), modifying its interface to a lesser or greater extent. It is very similar to a snit::widget. See WIDGET ADAPTORS.
You create an instance of a snit::type by passing the new instance's name to the type's create method. In the following example, we create a dog object called spot.
% snit::type dog {
# ....
}
::dog
% dog create spot
::spot
%
In general, the create method name can be omitted so long as the instance name doesn't conflict with any defined TYPE METHODS. (See TYPE COMPONENTS for the special case in which this doesn't work.) So the following example is identical to the previous example:
% snit::type dog {
# ....
}
::dog
% dog spot
::spot
%
This document generally uses the shorter form.
If the dog type defines OPTIONS, these can usually be given defaults at creation time:
% snit::type dog {
option -breed mongrel
option -color brown
method bark {} { return "$self barks." }
}
::dog
% dog create spot -breed dalmation -color spotted
::spot
% spot cget -breed
dalmation
% spot cget -color
spotted
%
Once created, the instance name now names a new Tcl command that is used to manipulate the object. For example, the following code makes the dog bark:
% spot bark ::spot barks. %
Some programmers prefer to save the object name in a variable, and reference it that way. For example,
% snit::type dog { ... }
::dog
% set d [dog spot -breed dalmation -color spotted]
::spot
% $d cget -breed
dalmation
% $d bark
::spot barks.
%
If you prefer this style, you might prefer to have Snit generate the instance's name automatically.
If you'd like Snit to generate an object name for you, use the %AUTO% keyword as the requested name:
% snit::type dog { ... }
::dog
% set d [dog %AUTO%]
::dog2
% $d bark
::dog2 barks.
%
The %AUTO% keyword can be embedded in a longer string:
% set d [dog obj_%AUTO%] ::obj_dog4 % $d bark ::obj_dog4 barks. %
Tcl's rename command renames other commands. It's a common technique in Tcl to modify an existing command by renaming it and defining a new command with the original name; the new command usually calls the renamed command.
snit::type commands, however, should never be renamed; to do so breaks the connection between the type and its objects.
Tcl's rename command renames other commands. It's a common technique in Tcl to modify an existing command by renaming it and defining a new command with the original name; the new command usually calls the renamed command.
All Snit objects (including widgets and widgetadaptors) can be renamed, though this flexibility has some consequences:
You'll get an error if .btn calls your command after your object is renamed.
.btn configure -command [list $self ButtonPress]
The mymethod command returns code that will call the desired method safely; the caller of the callback can add additional arguments to the end of the command as usual.
.btn configure -command [mymethod ButtonPress]
Any Snit object of any type can be destroyed by renaming it to the empty string using the Tcl rename command.
Snit megawidgets (i.e., instances of snit::widget and snit::widgetadaptor) can be destroyed like any other widget: by using the Tk destroy command on the widget or on one of its ancestors in the window hierarchy.
Every instance of a snit::type has a destroy method:
% snit::type dog { ... }
::dog
% dog spot
::spot
% spot bark
::spot barks.
% spot destroy
% spot barks
invalid command name "spot"
%
Finally, every Snit type has a type method called destroy; calling it destroys the type and all of its instances:
% snit::type dog { ... }
::dog
% dog spot
::spot
% spot bark
::spot barks.
% dog destroy
% spot bark
invalid command name "spot"
% dog fido
invalid command name "dog"
%
An instance method is a procedure associated with a specific object and called as a subcommand of the object's command. It is given free access to all of the object's type variables, instance variables, and so forth.
Instance methods are defined in the type definition using the method statement. Consider the following code that might be used to add dogs to a computer simulation:
% snit::type dog {
method bark {} {
return "$self barks."
}
method chase {thing} {
return "$self chases $thing."
}
}
::dog
%
A dog can bark, and it can chase things.
The method statement looks just like a normal Tcl proc, except that it appears in a snit::type definition. Notice that every instance method gets an implicit argument called self; this argument contains the object's name. (There's more on implicit method arguments below.)
The method name becomes a subcommand of the object. For example, let's put a simulated dog through its paces:
% dog spot ::spot % spot bark ::spot barks. % spot chase cat ::spot chases cat. %
If method A needs to call method B on the same object, it does so just as a client does: it calls method B as a subcommand of the object itself, using the object name stored in the implicit argument self.
Suppose, for example, that our dogs never chase anything without barking at them:
% snit::type dog {
method bark {} {
return "$self barks."
}
method chase {thing} {
return "$self chases $thing. [$self bark]"
}
}
::dog
% dog spot
::spot
% spot bark
::spot barks.
% spot chase cat
::spot chases cat. ::spot barks.
%
Not really, so long as you avoid the standard instance method names: configure, configurelist, cget, destroy, and info. Also, method names consisting of multiple words define hierarchical methods.
An object's methods are subcommands of the object's instance command. Hierarchical methods allow an object's methods to have subcommands of their own; and these can in turn have subcommands, and so on. This allows the programmer to define a tree-shaped command structure, such as is used by many of the Tk widgets--the subcommands of the Tk text widget's tag method are hierarchical methods.
Define methods whose names consist of multiple words. These words define the hierarchy implicitly. For example, the following code defines a tag method with subcommands cget and configure:
snit::widget mytext {
method {tag configure} {tag args} { ... }
method {tag cget} {tag option} {...}
}
Note that there is no explicit definition for the tag method; it is implicit in the definition of tag configure and tag cget. If you tried to define tag explicitly in this example, you'd get an error.
As subcommands of subcommands.
% mytext .text % .text tag configure redtext -foreground red -background black % .text tag cget redtext -foreground red %
It's often useful to define private methods, that is, instance methods intended to be called only by other methods of the same object.
Snit doesn't implement any access control on instance methods, so all methods are de facto public. Conventionally, though, the names of public methods begin with a lower-case letter, and the names of private methods begin with an upper-case letter.
For example, suppose our simulated dogs only bark in response to other stimuli; they never bark just for fun. So the bark method becomes Bark to indicate that it is private:
% snit::type dog {
# Private by convention: begins with uppercase letter.
method Bark {} {
return "$self barks."
}
method chase {thing} {
return "$self chases $thing. [$self Bark]"
}
}
::dog
% dog fido
::fido
% fido chase cat
::fido chases cat. ::fido barks.
%
Method argument lists are defined just like normal Tcl proc argument lists; in particular, they can include arguments with default values and the args argument.
However, every method also has a number of implicit arguments provided by Snit in addition to those explicitly defined. The names of these implicit arguments may not used to name explicit arguments.
The arguments implicitly passed to every method are type, selfns, win, and self.
The implicit argument type contains the fully qualified name of the object's type:
% snit::type thing {
method mytype {} {
return $type
}
}
::thing
% thing something
::something
% something mytype
::thing
%
The implicit argument self contains the object's fully qualified name.
If the object's command is renamed, then $self will change to match in subsequent calls. Thus, your code should not assume that $self is constant unless you know for sure that the object will never be renamed.
% snit::type thing {
method myself {} {
return $self
}
}
::thing
% thing mutt
::mutt
% mutt myself
::mutt
% rename mutt jeff
% jeff myself
::jeff
%
Each Snit object has a private namespace in which to store its INSTANCE VARIABLES and OPTIONS. The implicit argument selfns contains the name of this namespace; its value never changes, and is constant for the life of the object, even if the object's name changes:
% snit::type thing {
method myNameSpace {} {
return $selfns
}
}
::thing
% thing jeff
::jeff
% jeff myNameSpace
::thing::Snit_inst3
% rename jeff mutt
% mutt myNameSpace
::thing::Snit_inst3
%
The above example reveals how Snit names an instance's private namespace; however, you should not write code that depends on the specific naming convention, as it might change in future releases.
The implicit argument win is defined for all Snit methods, though it really makes sense only for those of WIDGETS and WIDGET ADAPTORS. $win is simply the original name of the object, whether it's been renamed or not. For widgets and widgetadaptors, it is also therefore the name of a Tk window.
When a snit::widgetadaptor is used to modify the interface of a widget or megawidget, it must rename the widget's original command and replace it with its own.
Thus, using win whenever the Tk window name is called for means that a snit::widget or snit::widgetadaptor can be adapted by a snit::widgetadaptor. See WIDGETS for more information.
It depends on the context.
Suppose in my application I have a dog object named fido, and I want fido to bark when a Tk button called .bark is pressed. In this case, I create the callback command in the usual way, using list:
button .bark -text "Bark!" -command [list fido bark]
In typical Tcl style, we use a callback to hook two independent components together. But suppose that the dog object has a graphical interface and owns the button itself? In this case, the dog must pass one of its own instance methods to the button it owns. The obvious thing to do is this:
% snit::widget dog {
constructor {args} {
#...
button $win.barkbtn -text "Bark!" -command [list $self bark]
#...
}
}
::dog
%
(Note that in this example, our dog becomes a snit::widget, because it has GUI behavior. See WIDGETS for more.) Thus, if we create a dog called .spot, it will create a Tk button called .spot.barkbtn; when pressed, the button will call $self bark.
Now, this will work--provided that .spot is never renamed to something else. But surely renaming widgets is abnormal? And so it is--unless .spot is the hull component of a snit::widgetadaptor. If it is, then it will be renamed, and .spot will become the name of the snit::widgetadaptor object. When the button is pressed, the command $self bark will be handled by the snit::widgetadaptor, which might or might not do the right thing.
There's a safer way to do it, and it looks like this:
% snit::widget dog {
constructor {args} {
#...
button $win.barkbtn -text "Bark!" -command [mymethod bark]
#...
}
}
::dog
%
The command mymethod takes any number of arguments, and can be used like list to build up a callback command; the only difference is that mymethod returns a form of the command that won't change even if the instance's name changes.
On the other hand, you might prefer to allow a widgetadaptor to override a method such that your renamed widget will call the widgetadaptor's method instead of its own. In this case, using [list $self bark] will do what you want...but this is a technique which should be used only in carefully controlled circumstances.
See DELEGATION.
An instance variable is a private variable associated with some particular Snit object. Instance variables can be scalars or arrays.
Scalar instance variables are defined in the type definition using the variable statement. You can simply name it, or you can initialize it with a value:
snit::type mytype {
# Define variable "greeting" and initialize it with "Howdy!"
variable greeting "Howdy!"
}
Array instance variables are also defined in the type definition using the variable command. You can initialize them at the same time by specifying the -array option:
snit::type mytype {
# Define array variable "greetings"
variable greetings -array {
formal "Good Evening"
casual "Howdy!"
}
}
Variables do not really exist until they are given values. If you do not initialize a variable when you define it, then you must be sure to assign a value to it (in the constructor, say, or in some method) before you reference it.
Just a few.
First, every Snit object has a built-in instance variable called options, which should never be redefined.
Second, all names beginning with "Snit_" are reserved for use by Snit internal code.
Third, instance variable names containing the namespace delimiter (::) are likely to cause great confusion.
No. Once you've defined an instance variable in the type definition, it can be used in any instance code (instance methods, the constructor, and the destructor) without declaration. This differs from normal Tcl practice, in which all non-local variables in a proc need to be declared.
There is a speed penalty to having all instance variables implicitly available in all instance code. Even though your code need not declare the variables explicitly, Snit must still declare them, and that takes time. If you have ten instance variables, a method that uses none of them must still pay the declaration penalty for all ten. In most cases, the additional runtime cost is negligible. If extreme cases, you might wish to avoid it; there are two methods for doing so.
The first is to define a single instance variable, an array, and store all of your instance data in the array. This way, you're only paying the declaration penalty for one variable--and you probably need the variable most of the time anyway. This method breaks down if your instance variables include multiple arrays; in Tcl 8.5, however, the dict command might come to your rescue.
The second method is to declare your instance variables explicitly in your instance code, while not including them in the type definition:
snit::type dog {
constructor {} {
variable mood
set mood happy
}
method setmood {newMood} {
variable mood
set mood $newMood
}
method getmood {} {
variable mood
return $mood
}
}
This allows you to ensure that only the required variables are included in each
method, at the cost of longer code and run-time errors when you forget to
declare a variable you need.
In Tk, it's common to pass a widget a variable name; for example, Tk label widgets have a -textvariable option which names the variable which will contain the widget's text. This allows the program to update the label's value just by assigning a new value to the variable.
If you naively pass the instance variable name to the label widget, you'll be confused by the result; Tk will assume that the name names a global variable. Instead, you need to provide a fully-qualified variable name. From within an instance method or a constructor, you can fully qualify the variable's name using the myvar command:
snit::widget mywidget {
variable labeltext ""
constructor {args} {
# ...
label $win.label -textvariable [myvar labeltext]
# ...
}
}
Practically speaking, you don't. Instead, you'll implement public variables as OPTIONS. Alternatively, you can write INSTANCE METHODS to set and get the variable's value.
A type's options are the equivalent of what other object-oriented languages would call public member variables or properties: they are data values which can be retrieved and (usually) set by the clients of an object.
Snit's implementation of options follows the Tk model fairly exactly, except that snit::type objects usually don't interact with THE TK OPTION DATABASE; snit::widget and snit::widgetadaptor objects, on the other hand, always do.
Options are defined in the type definition using the option statement. Consider the following type, to be used in an application that manages a list of dogs for a pet store:
snit::type dog {
option -breed -default mongrel
option -color -default brown
option -akc -default 0
option -shots -default 0
}
According to this, a dog has four notable properties: a breed, a color, a flag that says whether it's pedigreed with the American Kennel Club, and another flag that says whether it has had its shots. The default dog, evidently, is a brown mutt.
There are a number of options you can specify when defining an option; if -default is the only one, you can omit the word -default as follows:
snit::type dog {
option -breed mongrel
option -color brown
option -akc 0
option -shots 0
}
If no -default value is specified, the option's default value will be the empty string (but see THE TK OPTION DATABASE).
The Snit man page refers to options like these as "locally defined" options.
The normal convention is that the client may pass any number of options and their values after the object's name at object creation. For example, the ::dog command defined in the previous answer can now be used to create individual dogs. Any or all of the options may be set at creation time.
% dog spot -breed beagle -color "mottled" -akc 1 -shots 1 ::spot % dog fido -shots 1 ::fido %
So ::spot is a pedigreed beagle; ::fido is a typical mutt, but his owners evidently take care of him, because he's had his shots.
Note: If the type defines a constructor, it can specify a different object-creation syntax. See CONSTRUCTORS for more information.
Retrieve option values using the cget method:
% spot cget -color mottled % fido cget -breed mongrel %
Any number of options may be set at one time using the configure instance method. Suppose that closer inspection shows that ::fido is not a brown mongrel, but rather a rare Arctic Boar Hound of a lovely dun color:
% fido configure -color dun -breed "Arctic Boar Hound" % fido cget -color dun % fido cget -breed Arctic Boar Hound
Alternatively, the configurelist method takes a list of options and values; occasionally this is more convenient:
% set features [list -color dun -breed "Arctic Boar Hound"]
-color dun -breed {Arctic Boar Hound}
% fido configurelist $features
% fido cget -color
dun
% fido cget -breed
Arctic Boar Hound
%
In Tcl 8.5, the * keyword can be used with configure in this case:
% set features [list -color dun -breed "Arctic Boar Hound"]
-color dun -breed {Arctic Boar Hound}
% fido configure {*}$features
% fido cget -color
dun
% fido cget -breed
Arctic Boar Hound
%
The results are the same.
There are two ways an instance method can set and retrieve an option's value. One is to use the configure and cget methods, as shown below.
% snit::type dog {
option -weight 10
method gainWeight {} {
set wt [$self cget -weight]
incr wt
$self configure -weight $wt
}
}
::dog
% dog fido
::fido
% fido cget -weight
10
% fido gainWeight
% fido cget -weight
11
%
Alternatively, Snit provides a built-in array instance variable called options. The indices are the option names; the values are the option values. The method gainWeight can thus be rewritten as follows:
method gainWeight {} {
incr options(-weight)
}
As you can see, using the options variable involves considerably less typing and is the usual way to do it. But if you use -configuremethod or -cgetmethod (described in the following answers), you might wish to use the configure and cget methods anyway, just so that any special processing you've implemented is sure to get done. Also, if the option is delegated to a component then configure and cget are the only way to access it without accessing the component directly. See DELEGATION for more information.
Define the option with -readonly yes.
Suppose you've got an option that determines how instances of your type are constructed; it must be set at creation time, after which it's constant. For example, a dog never changes its breed; it might or might not have had its shots, and if not can have them at a later time. -breed should be read-only, but -shots should not be.
% snit::type dog {
option -breed -default mongrel -readonly yes
option -shots -default no
}
::dog
% dog fido -breed retriever
::fido
% fido configure -shots yes
% fido configure -breed terrier
option -breed can only be set at instance creation
%
Define a -cgetmethod for the option.
A -cgetmethod is a method that's called whenever the related option's value is queried via the cget instance method. The handler can compute the option's value, retrieve it from a database, or do anything else you'd like it to do.
Here's what the default behavior would look like if written using a -cgetmethod:
snit::type dog {
option -color -default brown -cgetmethod GetOption
method GetOption {option} {
return $options($option)
}
}
Any instance method can be used, provided that it takes one argument, the name of the option whose value is to be retrieved.
Define a -configuremethod for the option.
A -configuremethod is a method that's called whenever the related option is given a new value via the configure or configurelist instance methods. The method can pass the value on to some other object, store it in a database, or do anything else you'd like it to do.
Here's what the default configuration behavior would look like if written using a -configuremethod:
snit::type dog {
option -color -default brown -configuremethod SetOption
method SetOption {option value} {
set options($option) $value
}
}
Any instance method can be used, provided that it takes two arguments, the name of the option and the new value.
Note that if your method doesn't store the value in the options array, the options array won't get updated.
Define a -validatemethod.
A -validatemethod is a method that's called whenever the related option is given a new value via the configure or configurelist instance methods. It's the method's responsibility to determine whether the new value is valid, and throw an error if it isn't. The -validatemethod, if any, is called before the value is stored in the options array; in particular, it's called before the -configuremethod, if any.
For example, suppose an option always takes a Boolean value. You can ensure that the value is in fact a valid Boolean like this:
% snit::type dog {
option -shots -default no -validatemethod BooleanOption
method BooleanOption {option value} {
if {![string is boolean -strict $value]} {
error "expected a boolean value, got \"$value\""
}
}
}
::dog
% dog fido
% fido configure -shots yes
% fido configure -shots NotABooleanValue
expected a boolean value, got "NotABooleanValue"
%
Note that the same -validatemethod can be used to validate any number of
boolean options.
Any method can be a -validatemethod provided that it takes two arguments, the option name and the new option value.
A type variable is a private variable associated with a Snit type rather than with a particular instance of the type. In C++ and Java, the term static member variable is used for the same notion. Type variables can be scalars or arrays.
Scalar type variables are defined in the type definition using the typevariable statement. You can simply name it, or you can initialize it with a value:
snit::type mytype {
# Define variable "greeting" and initialize it with "Howdy!"
typevariable greeting "Howdy!"
}
Every object of type mytype now has access to a single variable called greeting.
Array-valued type variables are also defined using the typevariable command; to initialize them, include the -array option:
snit::type mytype {
# Define typearray variable "greetings"
typevariable greetings -array {
formal "Good Evening"
casual "Howdy!"
}
}
Variables do not really exist until they are given values. If you do not initialize a variable when you define it, then you must be sure to assign a value to it (in the type constructor, say) before you reference it.
Type variable names have the same restrictions as the names of INSTANCE VARIABLES do.
No. Once you've defined a type variable in the type definition, it can be used in INSTANCE METHODS or TYPE METHODS without declaration. This differs from normal Tcl practice, in which all non-local variables in a proc need to be declared.
Type variables are subject to the same speed/readability tradeoffs as instance variables; see Do I need to declare my instance variables in my methods?
In Tk, it's common to pass a widget a variable name; for example, Tk label widgets have a -textvariable option which names the variable which will contain the widget's text. This allows the program to update the label's value just by assigning a new value to the variable.
If you naively pass a type variable name to the label widget, you'll be confused by the result; Tk will assume that the name names a global variable. Instead, you need to provide a fully-qualified variable name. From within an instance method or a constructor, you can fully qualify the type variable's name using the mytypevar command:
snit::widget mywidget {
typevariable labeltext ""
constructor {args} {
# ...
label $win.label -textvariable [mytypevar labeltext]
# ...
}
}
There are two ways to do this. The preferred way is to write a pair of TYPE METHODS to set and query the type variable's value.
Type variables are stored in the type's namespace, which has the same name as the type itself. Thus, you can also publicize the type variable's name in your documentation so that clients can access it directly. For example,
snit::type mytype {
typevariable myvariable
}
set ::mytype::myvariable "New Value"
A type method is a procedure associated with the type itself rather than with any specific instance of the type, and called as a subcommand of the type command.
Type methods are defined in the type definition using the typemethod statement:
snit::type dog {
# List of pedigreed dogs
typevariable pedigreed
typemethod pedigreedDogs {} {
return $pedigreed
}
}
Suppose the dog type maintains a list of the names of the dogs that have pedigrees. The pedigreedDogs type method returns this list.
The typemethod statement looks just like a normal Tcl proc, except that it appears in a snit::type definition. Notice that every type method gets an implicit argument called type, which contains the fully-qualified type name.
The type method name becomes a subcommand of the type's command. For example, assuming that the constructor adds each pedigreed dog to the list of pedigreedDogs,
snit::type dog {
option -pedigreed 0
# List of pedigreed dogs
typevariable pedigreed
typemethod pedigreedDogs {} {
return $pedigreed
}
# ...
}
dog spot -pedigreed 1
dog fido
foreach dog [dog pedigreedDogs] { ... }
Not really, so long as you avoid the standard type method names: create, destroy, and info.
It's sometimes useful to define private type methods, that is, type methods intended to be called only by other type or instance methods of the same object.
Snit doesn't implement any access control on type methods; by convention, the names of public methods begin with a lower-case letter, and the names of private methods begin with an upper-case letter.
Alternatively, a Snit proc can be used as a private type method; see PROCS.
Method argument lists are defined just like normal Tcl proc argument lists; in particular, they can include arguments with default values and the args argument.
However, every type method is called with an implicit argument called type that contains the name of the type command. In addition, type methods should by convention avoid using the names of the arguments implicitly defined for INSTANCE METHODS.
If an instance or type method needs to call a type method, it should use $type to do so:
snit::type dog {
typemethod pedigreedDogs {} { ... }
typemethod printPedigrees {} {
foreach obj [$type pedigreedDogs] { ... }
}
}
It's common in Tcl to pass a snippet of code to another object, for it to call later. Because types cannot be renamed, you can just use the type name, or, if the callback is registered from within a type method, type. For example, suppose we want to print a list of pedigreed dogs when a Tk button is pushed:
button .btn -text "Pedigrees" -command [list dog printPedigrees] pack .btnAlternatively, from a method or type method you can use the mytypemethod command, just as you would use mymethod to define a callback command for INSTANCE METHODS.
Yes, you can define hierarchical type methods in just the same way as you can define hierarchical instance methods. See INSTANCE METHODS for more.
A Snit proc is really just a Tcl proc defined within the type's namespace. You can use procs for private code that isn't related to any particular instance.
Procs are defined by including a proc statement in the type definition:
snit::type mytype {
# Pops and returns the first item from the list stored in the
# listvar, updating the listvar
proc pop {listvar} { ... }
# ...
}
Any name can be used, so long as it does not begin with Snit_; names beginning with Snit_ are reserved for Snit's own use. However, the wise programmer will avoid proc names (set, list, if, etc.) that would shadow standard Tcl command names.
proc names, being private, should begin with a capital letter according to convention; however, as there are typically no public procs in the type's namespace it doesn't matter much either way.
Just like it calls any Tcl command. For example,
snit::type mytype {
# Pops and returns the first item from the list stored in the
# listvar, updating the listvar
proc pop {listvar} { ... }
variable requestQueue {}
# Get one request from the queue and process it.
method processRequest {} {
set req [pop requestQueue]
}
}
The myproc command returns a callback command for the proc, just as mymethod does for a method.
A type constructor is a body of code that initializes the type as a whole, rather like a C++ static initializer. The body of a type constructor is executed once when the type is defined, and never again.
A type can have at most one type constructor.
A type constructor is defined by using the typeconstructor statement in the type definition. For example, suppose the type uses an array-valued type variable as a look-up table, and the values in the array have to be computed at start-up.
% snit::type mytype {
typevariable lookupTable
typeconstructor {
array set lookupTable {key value...}
}
}
In object-oriented programming, an object's constructor is responsible for initializing the object completely at creation time. The constructor receives the list of options passed to the snit::type command's create method and can then do whatever it likes. That might include computing instance variable values, reading data from files, creating other objects, updating type and instance variables, and so forth.
The constructor's return value is ignored (unless it's an error, of course).
A constructor is defined by using the constructor statement in the type definition. Suppose that it's desired to keep a list of all pedigreed dogs. The list can be maintained in a type variable and retrieved by a type method. Whenever a dog is created, it can add itself to the list--provided that it's registered with the American Kennel Club.
% snit::type dog {
option -akc 0
typevariable akcList {}
constructor {args} {
$self configurelist $args
if {$options(-akc)} {
lappend akcList $self
}
}
typemethod akclist {} {
return $akcList
}
}
::dog
% dog spot -akc 1
::spot
% dog fido
::fido
% dog akclist
::spot
%
If you don't provide a constructor explicitly, you get the default constructor, which is identical to the explicitly-defined constructor shown here:
snit::type dog {
constructor {args} {
$self configurelist $args
}
}
When the constructor is called, args will be set to the list of arguments that follow the object's name. The constructor is allowed to interpret this list any way it chooses; the normal convention is to assume that it's a list of option names and values, as shown in the example above. If you simply want to save the option values, you should use the configurelist method, as shown.
Yes, you can. For example, suppose we wanted to be sure that the breed was explicitly stated for every dog at creation time, and couldn't be changed thereafter. One way to do that is as follows:
% snit::type dog {
variable breed
option -color brown
option -akc 0
constructor {theBreed args} {
set breed $theBreed
$self configurelist $args
}
method breed {} { return $breed }
}
::dog
% dog spot dalmatian -color spotted -akc 1
::spot
% spot breed
dalmatian
The drawback is that this syntax is non-standard, and may limit the compatibility of your new type with other people's code. For example, Snit assumes that it can create COMPONENTS using the standard creation syntax.
Constructor argument lists are subject to the same limitations as those on instance method argument lists. It has the same implicit arguments, and can contain default values and the args argument.
Yes. Writing the constructor can be tricky if you're delegating options to components, and there are specific issues relating to snit::widgets and snit::widgetadaptors. See DELEGATION, WIDGETS, WIDGET ADAPTORS, and THE TK OPTION DATABASE.
A destructor is a special kind of method that's called when an object is destroyed. It's responsible for doing any necessary clean-up when the object goes away: destroying COMPONENTS, closing files, and so forth.
Destructors are defined by using the destructor statement in the type definition.
Suppose we're maintaining a list of pedigreed dogs; then we'll want to remove dogs from it when they are destroyed.
snit::type dog {
option -akc 0
typevariable akcList {}
constructor {args} {
$self configurelist $args
if {$options(-akc)} {
lappend akcList $self
}
}
destructor {
set ndx [lsearch $akcList $self]
if {$ndx != -1} {
set akcList [lreplace $akcList $ndx $ndx]
}
}
typemethod akclist {} {
return $akcList
}
}
Yes; a destructor has no explicit arguments.
The destructor gets the same implicit arguments that are passed to INSTANCE METHODS: type, selfns, win, and self.
Yes and no.
Any Tk widgets created by a snit::widget or snit::widgetadaptor will be destroyed automatically by Tk when the megawidget is destroyed, in keeping with normal Tk behavior (destroying a parent widget destroys the whole tree).
Components of normal snit::types, on the other hand, are never destroyed automatically, nor are non-widget components of Snit megawidgets. If your object creates them in its constructor, then it should generally destroy them in its destructor.
Yes. If an object's constructor throws an error, the object's destructor will be called to clean up; this means that the object might not be completely constructed when the destructor is called. This can cause the destructor to throw its own error; the result is usually misleading, confusing, and unhelpful. Consequently, it's important to write your destructor so that it's fail-safe.
For example, a dog might create a tail component; the component will need to be destroyed. But suppose there's an error while processing the creation options--the destructor will be called, and there will be no tail to destroy. The simplest solution is generally to catch and ignore any errors while destroying components.
snit::type dog {
component tail
constructor {args} {
$self configurelist $args
set tail [tail %AUTO%]
}
destructor {
catch {$tail destroy}
}
}
Often an object will create and manage a number of other objects. A Snit megawidget, for example, will often create a number of Tk widgets. These objects are part of the main object; it is composed of them, so they are called components of the object.
But Snit also has a more precise meaning for COMPONENT. The components of a Snit object are those objects to which methods or options can be delegated. (See DELEGATION for more information about delegation.)
First, you must decide what role a component plays within your object, and give the role a name. Then, you declare the component using its role name and the component statement. The component statement declares an instance variable which is used to store the component's command name when the component is created.
For example, suppose your dog object creates a tail object (the better to wag with, no doubt):
snit::type dog {
component mytail
constructor {args} {
# Create and save the component's command
set mytail [tail %AUTO% -partof $self]
$self configurelist $args
}
method wag {} {
$mytail wag
}
}
As shown here, it doesn't matter what the tail object's real name is; the dog object refers to it by its component name.
The above example shows one way to delegate the wag method to the mytail component; see DELEGATION for an easier way.
A component has two names. The first name is that of the component variable; this represents the role the component object plays within the Snit object. This is the component name proper, and is the name used to refer to the component within Snit code. The second name is the name of the actual component object created by the Snit object's constructor. This second name is always a Tcl command name, and is referred to as the component's object name.
In the example in the previous question, the component name is mytail; the mytail component's object name is chosen automatically by Snit since %AUTO% was used when the component object was created.
Yes. snit::widget and snit::widgetadaptor objects have a special component called the hull component; thus, the name hull should be used for no other purpose.
Otherwise, since component names are in fact instance variable names they must follow the rules for INSTANCE VARIABLES.
An owned component is a component whose object command's lifetime is controlled by the snit::type or snit::widget.
As stated above, a component is an object to which our object can delegate methods or options. Under this definition, our object will usually create its component objects, but not necessarily. Consider the following: a dog object has a tail component; but tail knows that it's part of the dog:
snit::type dog {
component mytail
constructor {args} {
set mytail [tail %AUTO% -partof $self]
$self configurelist $args
}
destructor {
catch {$mytail destroy}
}
delegate method wagtail to mytail as wag
method bark {} {
return "$self barked."
}
}
snit::type tail {
component mydog
option -partof -readonly yes
constructor {args} {
$self configurelist $args
set mydog $options(-partof)
}
method wag {} {
return "Wag, wag."
}
method pull {} {
$mydog bark
}
}
Thus, if you ask a dog to wag its tail, it tells its tail to wag; and if you
pull the dog's tail, the tail tells the dog to bark. In this scenario, the
tail is a component of the dog, and the dog is a component of the tail, but
the dog owns the tail and not the other way around.
The install command creates an owned component using a specified command, and assigns the result to the component's instance variable. For example:
snit::type dog {
component mytail
constructor {args} {
# set mytail [tail %AUTO% -partof $self]
install mytail using tail %AUTO% -partof $self
$self configurelist $args
}
}
In a snit::type's code, the install command shown above is equivalent to the set mytail command that's commented out. In a snit::widget's or snit::widgetadaptor's, code, however, the install command also queries THE TK OPTION DATABASE and initializes the new component's options accordingly. For consistency, it's a good idea to get in the habit of using install for all owned components.
No, not necessarily. In fact, there's no reason why an object can't destroy and recreate a component multiple times over its own lifetime.
Yes.
Component objects which are Tk widgets or megawidgets must have valid Tk window names.
Component objects which are not widgets or megawidgets must have fully-qualified command names, i.e., names which include the full namespace of the command. Note that Snit always creates objects with fully qualified names.
Next, the object names of components and owned by your object must be unique. This is no problem for widget components, since widget names are always unique; but consider the following code:
snit::type tail { ... }
snit::type dog {
delegate method wag to mytail
constructor {} {
install mytail using tail mytail
}
}
This code uses the component name, mytail, as the component object name. This is not good, and here's why: Snit instance code executes in the Snit type's namespace. In this case, the mytail component is created in the ::dog:: namespace, and will thus have the name ::dog::mytail.
Now, suppose you create two dogs. Both dogs will attempt to create a tail called ::dog::mytail. The first will succeed, and the second will fail, since Snit won't let you create an object if its name is already a command. Here are two ways to avoid this situation:
First, if the component type is a snit::type you can specify %AUTO% as its name, and be guaranteed to get a unique name. This is the safest thing to do:
install mytail using tail %AUTO%
If the component type isn't a snit::type you can create the component in the object's instance namespace:
install mytail using tail ${selfns}::mytail
Make sure you pick a unique name within the instance namespace.
That depends. When a parent widget is destroyed, all child widgets are destroyed automatically. Thus, if your object is a snit::widget or snit::widgetadaptor you don't need to destroy any components that are widgets, because they will generally be children or descendants of your megawidget.
If your object is an instance of snit::type, though, none of its owned components will be destroyed automatically, nor will be non-widget components of a snit::widget be destroyed automatically. All such owned components must be destroyed explicitly, or they won't be destroyed at all.
Yes, and there are two ways to do it. The most appropriate way is usually to use DELEGATION. Delegation allows you to pass the options and methods you specify along to particular components. This effectively hides the components from the users of your type, and ensures good encapsulation.
However, there are times when it's appropriate, not to mention simpler, just to make the entire component part of your type's public interface.
When you declare the component, specify the component statement's -public option. The value of this option is the name of a method which will be delegated to your component's object command.
For example, supposed you've written a combobox megawidget which owns a listbox widget, and you want to make the listbox's entire interface public. You can do it like this:
snit::widget combobox {
component listbox -public listbox
constructor {args} {
install listbox using listbox $win.listbox ....
}
}
combobox .mycombo
Your comobox widget, .mycombo, now has a listbox method which has all of the same subcommands as the listbox widget itself. Thus, the above code sets the listbox component's width to 30.
Usually you'll let the method name be the same as the component name; however, you can name it anything you like.
A type component is a component that belongs to the type itself instead of to a particular instance of the type. The relationship between components and type components is the same as the relationship between INSTANCE VARIABLES and TYPE VARIABLES. Both INSTANCE METHODS and TYPE METHODS can be delegated to type components.
Once you understand COMPONENTS and DELEGATION, type components are just more of the same.
Declare a type component using the typecomponent statement. It takes the same options (-inherit and -public) as the component statement does, and defines a type variable to hold the type component's object command.
Suppose in your model you've got many dogs, but only one veterinarian. You might make the veterinarian a type component.
snit::type veterinarian { ... }
snit::type dog {
typecomponent vet
# ...
}
Just use the set command to assign the component's object command to the type component. Because types (even snit::widget types) are not widgets, and do not have options anyway, the extra features of the install command are not needed.
You'll usually install type components in the type constructor, as shown here:
snit::type veterinarian { ... }
snit::type dog {
typecomponent vet
typeconstructor {
set vet [veterinarian %AUTO%]
}
}
Yes, the same as on INSTANCE VARIABLES, TYPE VARIABLES, and normal COMPONENTS.
Delegation, simply put, is when you pass a task you've been given to one of your assistants. (You do have assistants, don't you?) Snit objects can do the same thing. The following example shows one way in which the dog object can delegate its wag method and its -taillength option to its tail component.
snit::type dog {
variable mytail
option -taillength -configuremethod SetTailOption -cgetmethod GetTailOption
method SetTailOption {option value} {
$mytail configure $option $value
}
method GetTailOption {option} {
$mytail cget $option
}
method wag {} {
$mytail wag
}
constructor {args} {
install mytail using tail %AUTO% -partof $self
$self configurelist $args
}
}
This is the hard way to do it, by it demonstrates what delegation is all about. See the following answers for the easy way to do it.
Note that the constructor calls the configurelist method after it creates its tail; otherwise, if -taillength appeared in the list of args we'd get an error.
Delegation occurs frequently enough that Snit makes it easy. Any method can be delegated to any component or type component by placing a single delegate statement in the type definition. (See COMPONENTS and TYPE COMPONENTS for more information about component names.)
For example, here's a much better way to delegate the dog object's wag method:
% snit::type dog {
delegate method wag to mytail
constructor {} {
install mytail using tail %AUTO%
}
}
::dog
% snit::type tail {
method wag {} { return "Wag, wag, wag."}
}
::tail
% dog spot
::spot
% spot wag
Wag, wag, wag.
This code has the same effect as the code shown under the previous question: when a dog's wag method is called, the call and its arguments are passed along automatically to the tail object.
Note that when a component is mentioned in a delegate statement, the component's instance variable is defined implicitly. However, it's still good practice to declare it explicitly using the component statement.
Note also that you can define a method name using the method statement, or you can define it using delegate; you can't do both.
Suppose you wanted to delegate the dog's wagtail method to the tail's wag method. After all you wag the tail, not the dog. It's easily done:
snit::type dog {
delegate method wagtail to mytail as wag
constructor {args} {
install mytail using tail %AUTO% -partof $self
$self configurelist $args
}
}
Suppose the tail's wag method takes as an argument the number of times the tail should be wagged. You want to delegate the dog's wagtail method to the tail's wag method, specifying that the tail should be wagged exactly three times. This is easily done, too:
snit::type dog {
delegate method wagtail to mytail as {wag 3}
# ...
}
snit::type tail {
method wag {count} {
return [string repeat "Wag " $count]
}
# ...
}
Normal method delegation assumes that you're delegating a method (a subcommand of an object command) to a method of another object (a subcommand of a different object command). But not all Tcl objects follow Tk conventions, and not everything you'd to which you'd like to delegate a method is necessary an object. Consequently, Snit makes it easy to delegate a method to pretty much anything you like using the delegate statement's using clause.
Suppose your dog simulation stores dogs in a database, each dog as a single record. The database API you're using provides a number of commands to manage records; each takes the record ID (a string you choose) as its first argument. For example, saverec saves a record. If you let the record ID be the name of the dog object, you can delegate the dog's save method to the saverec command as follows:
snit::type dog {
delegate method save using {saverec %s}
}
The %s is replaced with the instance name when the save method is
called; any additional arguments are the appended to the resulting command.
The using clause understands a number of other %-conversions; in addition to the instance name, you can substitute in the method name (%m), the type name (%t), the instance namespace (%n), the Tk window name (%w), and, if a component or typecomponent name was given in the delegate statement, the component's object command (%c).
Just exactly as you would to a component object. The delegate method statement accepts both component and type component names in its to clause.
Use the delegate typemethod statement. It works like delegate method, with these differences: first, it defines a type method instead of an instance method; second, the using clause ignores the %s, %n, and %w %-conversions.
Naturally, you can't delegate a type method to an instance component...Snit wouldn't know which instance should receive it.
The first question in this section (see DELEGATION) shows one way to delegate an option to a component; but this pattern occurs often enough that Snit makes it easy. For example, every tail object has a -length option; we want to allow the creator of a dog object to set the tail's length. We can do this:
% snit::type dog {
delegate option -length to mytail
constructor {args} {
install mytail using tail %AUTO% -partof $self
$self configurelist $args
}
}
::dog
% snit::type tail {
option -partof
option -length 5
}
::tail
% dog spot -length 7
::spot
% spot cget -length
7
This produces nearly the same result as the -configuremethod and -cgetmethod shown under the first question in this section: whenever a dog object's -length option is set or retrieved, the underlying tail object's option is set or retrieved in turn.
Note that you can define an option name using the option statement, or you can define it using delegate; you can't do both.
In the previous answer we delegated the dog's -length option down to its tail. This is, of course, wrong. The dog has a length, and the tail has a length, and they are different. What we'd really like to do is give the dog a -taillength option, but delegate it to the tail's -length option:
snit::type dog {
delegate option -taillength to mytail as -length
constructor {args} {
set mytail [tail %AUTO% -partof $self]
$self configurelist $args
}
}
It may happen that a Snit object gets most of its behavior from one of its components. This often happens with snit::widgetadaptors, for example, where we wish to slightly the modify the behavior of an existing widget. To carry on with our dog example, however, suppose that we have a snit::type called animal that implements a variety of animal behaviors--moving, eating, sleeping, and so forth. We want our dog objects to inherit these same behaviors, while adding dog-like behaviors of its own. Here's how we can give a dog methods and options of its own while delegating all other methods and options to its animal component:
snit::type dog {
delegate option * to animal
delegate method * to animal
option -akc 0
constructor {args} {
install animal using animal %AUTO% -name $self
$self configurelist $args
}
method wag {} {
return "$self wags its tail"
}
}
That's it. A dog is now an animal that has a -akc option and can wag its tail.
Note that we don't need to specify the full list of method names or option names that animal will receive. It gets anything dog doesn't recognize--and if it doesn't recognize it either, it will simply throw an error, just as it should.
You can also delegate all unknown type methods to a type component using delegate typemethod *.
In the previous answer, we said that every dog is an animal by delegating all unknown methods and options to the animal component. But what if the animal type has some methods or options that we'd like to suppress?
One solution is to explicitly delegate all the options and methods, and forgo the convenience of delegate method * and delegate option *. But if we wish to suppress only a few options or methods, there's an easier way:
snit::type dog {
delegate option * to animal except -numlegs
delegate method * to animal except {fly climb}
# ...
constructor {args} {
install animal using animal %AUTO% -name $self -numlegs 4
$self configurelist $args
}
# ...
}
Dogs have four legs, so we specify that explicitly when we create the animal component, and explicitly exclude -numlegs from the set of delegated options. Similarly, dogs can neither fly nor climb, so we exclude those animal methods as shown.
Yes; just specify multiple words in the delegated method's name:
snit::type tail {
method wag {} {return "Wag, wag"}
method droop {} {return "Droop, droop"}
}
snit::type dog {
delegate method {tail wag} to mytail
delegate method {tail droop} to mytail
# ...
constructor {args} {
install mytail using tail %AUTO%
$self configurelist $args
}
# ...
}
Unrecognized hierarchical methods can also be delegated; the following code delegates all subcommands of the "tail" method to the "mytail" component:
snit::type dog {
delegate method {tail *} to mytail
# ...
}
A snit::widget is the Snit version of what Tcl programmers usually call a megawidget: a widget-like object usually consisting of one or more Tk widgets all contained within a Tk frame.
A snit::widget is also a special kind of snit::type. Just about everything in this FAQ list that relates to snit::types also applies to snit::widgets.
snit::widgets are defined using the snit::widget command, just as snit::types are defined by the snit::type command.
The body of the definition can contain all of the same kinds of statements, plus a couple of others which will be mentioned below.
Snit can't create a Tk widget object; only Tk can do that. Thus, every instance of a snit::widget must be wrapped around a genuine Tk widget; this Tk widget is called the hull component. Snit effectively piggybacks the behavior you define (methods, options, and so forth) on top of the hull component so that the whole thing behaves like a standard Tk widget.
For snit::widgets the hull component must be a Tk widget that defines the -class option.
snit::widgetadaptors differ from snit::widgets chiefly in that any kind of widget can be used as the hull component; see WIDGET ADAPTORS.
A snit::widget's hull component will usually be a Tk frame widget; however, it may be any Tk widget that defines the -class option. You can explicitly choose the hull type you prefer by including the hulltype command in the widget definition:
snit::widget mytoplevel {
hulltype toplevel
# ...
}
If no hulltype command appears, the hull will be a frame.
By default, Snit recognizes the following hull types: the Tk widgets frame, labelframe, toplevel, and the Tile widgets ttk::frame, ttk::labelframe, and ttk::toplevel. To enable the use of some other kind of widget as the hull type, you can lappend the widget command to the variable snit::hulltypes (always provided the widget defines the -class option. For example, suppose Tk gets a new widget type called a prettyframe:
lappend snit::hulltypes prettyframe
snit::widget mywidget {
hulltype prettyframe
# ...
}
Every widget, whether a genuine Tk widget or a Snit megawidget, has to have a valid Tk window name. When a snit::widget is first created, its instance name, self, is a Tk window name; however, if the snit::widget is used as the hull component by a snit::widgetadaptor its instance name will be changed to something else. For this reason, every snit::widget method, constructor, destructor, and so forth is passed another implicit argument, win, which is the window name of the megawidget. Any children should be named using win as the root.
Thus, suppose you're writing a toolbar widget, a frame consisting of a number of buttons placed side-by-side. It might look something like this:
snit::widget toolbar {
delegate option * to hull
constructor {args} {
button $win.open -text Open -command [mymethod open]
button $win.save -text Save -command [mymethod save]
# ....
$self configurelist $args
}
}
See also the question on renaming objects, toward the top of this file.
A snit::widgetadaptor is a kind of snit::widget. Whereas a snit::widget's hull is automatically created and is always a Tk frame, a snit::widgetadaptor can be based on any Tk widget--or on any Snit megawidget, or even (with luck) on megawidgets defined using some other package.
It's called a widget adaptor because it allows you to take an existing widget and customize its behavior.
Use the snit::widgetadaptor command. The definition for a snit::widgetadaptor looks just like that for a snit::type or snit::widget, except that the constructor must create and install the hull component.
For example, the following code creates a read-only text widget by the simple device of turning its insert and delete methods into no-ops. Then, we define new methods, ins and del, which get delegated to the hull component as insert and delete. Thus, we've adapted the text widget and given it new behavior while still leaving it fundamentally a text widget.
::snit::widgetadaptor rotext {
constructor {args} {
# Create the text widget; turn off its insert cursor
installhull using text -insertwidth 0
# Apply any options passed at creation time.
$self configurelist $args
}
# Disable the text widget's insert and delete methods, to
# make this readonly.
method insert {args} {}
method delete {args} {}
# Enable ins and del as synonyms, so the program can insert and
# delete.
delegate method ins to hull as insert
delegate method del to hull as delete
# Pass all other methods and options to the real text widget, so
# that the remaining behavior is as expected.
delegate method * to hull
delegate option * to hull
}
The most important part is in the constructor. Whereas snit::widget creates the hull for you, snit::widgetadaptor cannot -- it doesn't know what kind of widget you want. So the first thing the constructor does is create the hull component (a Tk text widget in this case), and then installs it using the installhull command.
Note: There is no instance command until you create one by installing a hull component. Any attempt to pass methods to $self prior to calling installhull will fail.
Yes.
At times, it can be convenient to adapt a pre-existing widget instead of creating your own. For example, the Bwidget PagesManager widget manages a set of frame widgets, only one of which is visible at a time. The application chooses which frame is visible. All of the These frames are created by the PagesManager itself, using its add method. It's convenient to adapt these frames to do what we'd like them to do.
In a case like this, the Tk widget will already exist when the snit::widgetadaptor is created. Snit provides an alternate form of the installhull command for this purpose:
snit::widgetadaptor pageadaptor {
constructor {args} {
# The widget already exists; just install it.
installhull $win
# ...
}
}
Maybe. If the other megawidget is a snit::widget or snit::widgetadaptor, then yes. If it isn't then, again, maybe. You'll have to try it and see. You're most likely to have trouble with widget destruction--you have to make sure that your megawidget code receives the <Destroy> event before the megawidget you're adapting does.
The Tk option database is a database of default option values maintained by Tk itself; every Tk application has one. The concept of the option database derives from something called the X Windows resource database; however, the option database is available in every Tk implementation, including those which do not use the X Windows system (e.g., Microsoft Windows).
Full details about the Tk option database are beyond the scope of this document; both Practical Programming in Tcl and Tk by Welch, Jones, and Hobbs, and Effective Tcl/Tk Programming by Harrison and McClennan., have good introductions to it.
Snit is implemented so that most of the time it will simply do the right thing with respect to the option database, provided that the widget developer does the right thing by Snit. The body of this section goes into great deal about what Snit requires. The following is a brief statement of the requirements, for reference.
The interaction of Tk widgets with the option database is a complex thing; the interaction of Snit with the option database is even more so, and repays attention to detail.
No, they don't; querying the option database requires a Tk window name, and snit::types don't have one.
If you create an instance of a snit::type as a component of a snit::widget or snit::widgetadaptor, on the other hand, and if any options are delegated to the component, and if you use install to create and install it, then the megawidget will query the option database on the snit::type's behalf. This might or might not be what you want, so take care.
Every Tk widget has a "widget class": a name that is used when adding option settings to the database. For Tk widgets, the widget class is the same as the widget command name with an initial capital. For example, the widget class of the Tk button widget is Button.
Similarly, the widget class of a snit::widget defaults to the unqualified type name with the first letter capitalized. For example, the widget class of
snit::widget ::mylibrary::scrolledText { ... }
is ScrolledText.
The widget class can also be set explicitly using the widgetclass statement within the snit::widget definition:
snit::widget ::mylibrary::scrolledText {
widgetclass Text
# ...
}
The above definition says that a scrolledText megawidget has the same widget class as an ordinary text widget. This might or might not be a good idea, depending on how the rest of the megawidget is defined, and how its options are delegated.
The widget class of a snit::widgetadaptor is just the widget class of its hull widget; Snit has no control over this.
Note that the widget class can be changed only for frame and toplevel widgets, which is why these are the valid hull types for snit::widgets.
Try to use snit::widgetadaptors only to make small modifications to another widget's behavior. Then, it will usually not make sense to change the widget's widget class anyway.
Every Tk widget option has three names: the option name, the resource name, and the class name. The option name begins with a hyphen and is all lowercase; it's used when creating widgets, and with the configure and cget commands.
The resource and class names are used to initialize option default values by querying the option database. The resource name is usually just the option name minus the hyphen, but may contain uppercase letters at word boundaries; the class name is usually just the resource name with an initial capital, but not always. For example, here are the option, resource, and class names for several Tk text widget options:
-background background Background
-borderwidth borderWidth BorderWidth
-insertborderwidth insertBorderWidth BorderWidth
-padx padX Pad
As is easily seen, sometimes the resource and class names can be inferred from the option name, but not always.
For options implicitly delegated to a component using delegate option *, the resource and class names will be exactly those defined by the component. The configure method returns these names, along with the option's default and current values:
% snit::widget mytext {
delegate option * to text
constructor {args} {
install text using text .text
# ...
}
# ...
}
::mytext
% mytext .text
% .text configure -padx
-padx padX Pad 1 1
%
For all other options (whether locally defined or explicitly delegated), the resource and class names can be defined explicitly, or they can be allowed to have default values.
By default, the resource name is just the option name minus the hyphen; the the class name is just the option name with an initial capital letter. For example, suppose we explicitly delegate "-padx":
% snit::widget mytext {
option -myvalue 5
delegate option -padx to text
delegate option * to text
constructor {args} {
install text using text .text
# ...
}
# ...
}
::mytext
% mytext .text
% .text configure -myvalue
-myvalue myvalue Myvalue 5 5
% .text configure -padx
-padx padx Padx 1 1
%
Here the resource and class names are chosen using the default rules. Often these rules are sufficient, but in the case of "-padx" we'd most likely prefer that the option's resource and class names are the same as for the built-in Tk widgets. This is easily done:
% snit::widget mytext {
delegate option {-padx padX Pad} to text
# ...
}
::mytext
% mytext .text
% .text configure -padx
-padx padX Pad 1 1
%
The option database is queried for each of the megawidget's locally-defined options, using the option's resource and class name. If the result isn't "", then it replaces the default value given in widget definition. In either case, the default can be overridden by the caller. For example,
option add *Mywidget.texture pebbled
snit::widget mywidget {
option -texture smooth
# ...
}
mywidget .mywidget -texture greasy
Here, -texture would normally default to "smooth", but because of the entry added to the option database it defaults to "pebbled". However, the caller has explicitly overridden the default, and so the new widget will be "greasy".
That depends on whether the options are delegated to the hull, or to some other component.
A snit::widget's hull is a widget, and given that its class has been set it is expected to query the option database for itself. The only exception concerns options that are delegated to it with a different name. Consider the following code:
option add *Mywidget.borderWidth 5
option add *Mywidget.relief sunken
option add *Mywidget.hullbackground red
option add *Mywidget.background green
snit::widget mywidget {
delegate option -borderwidth to hull
delegate option -hullbackground to hull as -background
delegate option * to hull
# ...
}
mywidget .mywidget
set A [.mywidget cget -relief]
set B [.mywidget cget -hullbackground]
set C [.mywidget cget -background]
set D [.mywidget cget -borderwidth]
The question is, what are the values of variables A, B, C and D?
The value of A is "sunken". The hull is a Tk frame which has been given the widget class Mywidget; it will automatically query the option database and pick up this value. Since the -relief option is implicitly delegated to the hull, Snit takes no action.
The value of B is "red". The hull will automatically pick up the value "green" for its -background option, just as it picked up the -relief value. However, Snit knows that -hullbackground is mapped to the hull's -background option; hence, it queries the option database for -hullbackground and gets "red" and updates the hull accordingly.
The value of C is also "red", because -background is implicitly delegated to the hull; thus, retrieving it is the same as retrieving -hullbackground. Note that this case is unusual; the -background option should probably have been excluded using the delegate statement's except clause, or (more likely) delegated to some other component.
The value of D is "5", but not for the reason you think. Note that as it is defined above, the resource name for -borderwidth defaults to borderwidth, whereas the option database entry is borderWidth, in accordance with the standard Tk naming for this option. As with -relief, the hull picks up its own -borderwidth option before Snit does anything. Because the option is delegated under its own name, Snit assumes that the correct thing has happened, and doesn't worry about it any further. To avoid confusion, the -borderwidth option should have been delegated like this:
delegate option {-borderwidth borderWidth BorderWidth} to hull
For snit::widgetadaptors, the case is somewhat altered. Widget adaptors retain the widget class of their hull, and the hull is not created automatically by Snit. Instead, the snit::widgetadaptor must call installhull in its constructor. The normal way to do this is as follows:
snit::widgetadaptor mywidget {
# ...
constructor {args} {
# ...
installhull using text -foreground white
# ...
}
# ...
}
In this case, the installhull command will create the hull using a command like this:
set hull [text $win -foreground white]
The hull is a text widget, so its widget class is Text. Just as with snit::widget hulls, Snit assumes that it will pick up all of its normal option values automatically, without help from Snit. Options delegated from a different name are initialized from the option database in the same way as described above.
In earlier versions of Snit, snit::widgetadaptors were expected to call installhull like this:
installhull [text $win -foreground white]
This form still works--but Snit will not query the option database as described above.
For hull components, Snit assumes that Tk will do most of the work automatically. Non-hull components are somewhat more complicated, because they are matched against the option database twice.
A component widget remains a widget still, and is therefore initialized from the option database in the usual way. A text widget remains a text widget whether it is a component of a megawidget or not, and will be created as such.
But then, the option database is queried for all options delegated to the component, and the component is initialized accordingly--provided that the install command is used to create it.
Before option database support was added to Snit, the usual way to create a component was to simply create it in the constructor and assign its command name to the component variable:
snit::widget mywidget {
delegate option -background to myComp
constructor {args} {
set myComp [text $win.text -foreground black]
}
}
The drawback of this method is that Snit has no opportunity to initialize the component properly. Hence, the following approach is now used:
snit::widget mywidget {
delegate option -background to myComp
constructor {args} {
install myComp using text $win.text -foreground black
}
}
The install command does the following:
You don't really need to know all of this; just use install to install your components, and Snit will try to do the right thing.
A snit::type never queries the option database. However, a snit::widget can have non-widget components. And if options are delegated to those components, and if the install command is used to install those components, then they will be initialized from the option database just as widget components are.
However, when used within a megawidget, install assumes that the created component uses a reasonably standard widget-like creation syntax. If it doesn't, don't use install.
An ensemble command is a command with subcommands. Snit objects are all ensemble commands; however, the term more usually refers to commands like the standard Tcl commands string, file, and clock. In a sense, these are singleton objects--there's only one instance of them.
There are two ways--as a snit::type, or as an instance of a snit::type.
Define a type whose INSTANCE METHODS are the subcommands of your ensemble command. Then, create an instance of the type with the desired name.
For example, the following code uses DELEGATION to create a work-alike for the standard string command:
snit::type ::mynamespace::mystringtype {
delegate method * to stringhandler
constructor {} {
set stringhandler string
}
}
::mynamespace::mystringtype mystring
We create the type in a namespace, so that the type command is hidden; then we
create a single instance with the desired name-- mystring, in this
case.
This method has two drawbacks. First, it leaves the type command floating about. More seriously, your shiny new ensemble command will have info and destroy subcommands that you probably have no use for. But read on.
Define a type whose TYPE METHODS are the subcommands of your ensemble command.
For example, the following code uses DELEGATION to create a work-alike for the standard string command:
snit::type mystring {
delegate typemethod * to stringhandler
typeconstructor {
set stringhandler string
}
}
Now the type command itself is your ensemble command.
This method has only one drawback, and though it's major, it's also surmountable. Your new ensemble command will have create, info and destroy subcommands you don't want. And worse yet, since the create method can be implicit, users of your command will accidentally be creating instances of your mystring type if they should mispell one of the subcommands. The command will succeed--the first time--but won't do what's wanted. This is very bad.
The work around is to set some PRAGMAS, as shown here:
snit::type mystring {
pragma -hastypeinfo no
pragma -hastypedestroy no
pragma -hasinstances no
delegate typemethod * to stringhandler
typeconstructor {
set stringhandler string
}
}
Here we've used the pragma statement to tell Snit that we don't want the
info typemethod or the destroy typemethod, and that our type has
no instances; this eliminates the create typemethod and all related
code. As a result, our ensemble command will be well-behaved, with no
unexpected subcommands.
A pragma is an option you can set in your type definitions that affects how the type is defined and how it works once it is defined.
Use the pragma statement. Each pragma is an option with a value; each time you use the pragma statement you can set one or more of them.
Set the -hastypeinfo pragma to no:
snit::type dog {
pragma -hastypeinfo no
# ...
}
Snit will refrain from defining the info type method.
Set the -hastypedestroy pragma to no:
snit::type dog {
pragma -hastypedestroy no
# ...
}
Snit will refrain from defining the destroy type method.
Set the -hasinstances pragma to no:
snit::type dog {
pragma -hasinstances no
# ...
}
Snit will refrain from defining the create type method; if you call the type command with an unknown method name, you'll get an error instead of a new instance of the type.
This is useful if you wish to use a snit::type to define an ensemble command rather than a type with instances.
Pragmas -hastypemethods and -hasinstances cannot both be false (or there'd be nothing left).
Normal Tk widget type commands don't have subcommands; all they do is create widgets--in Snit terms, the type command calls the create type method directly. To get the same behavior from Snit, set the -hastypemethods pragma to no:
snit::type dog {
pragma -hastypemethods no
#...
}
# Creates ::spot
dog spot
# Tries to create an instance called ::create
dog create spot
Pragmas -hastypemethods and -hasinstances cannot both be false (or there'd be nothing left).
Up until Snit 0.95, you could use any name for an instance of a snit::type, even if the name was already in use by some other object or command. You could do the following, for example:
snit::type dog { ... }
dog proc
You now have a new dog named "proc", which is probably not something that you really wanted to do. As a result, Snit now throws an error if your chosen instance name names an existing command. To restore the old behavior, set the -canreplace pragma to yes:
snit::type dog {
pragma -canreplace yes
# ...
}
In Snit 1.x, you can set the -simpledispatch pragma to yes.
Snit 1.x method dispatch is both flexible and fast, but the flexibility comes with a price. If your type doesn't require the flexibility, the -simpledispatch pragma allows you to substitute a simpler dispatch mechanism that runs quite a bit faster. The limitations are these:
In Snit 2.2, the -simpledispatch macro is obsolete, and ignored; all Snit 2.2 method dispatch is faster than Snit 1.x's -simpledispatch.
A Snit macro is nothing more than a Tcl proc that's defined in the Tcl interpreter used to compile Snit type definitions.
You can use Snit macros to define new type definition syntax, and to support conditional compilation.
Suppose you want your type to use a fast C extension if it's available; otherwise, you'll fallback to a slower Tcl implementation. You want to define one set of methods in the first case, and another set in the second case. But how can your type definition know whether the fast C extension is available or not?
It's easily done. Outside of any type definition, define a macro that returns 1 if the extension is available, and 0 otherwise:
if {$gotFastExtension} {
snit::macro fastcode {} {return 1}
} else {
snit::macro fastcode {} {return 0}
}
Then, use your macro in your type definition:
snit::type dog {
if {[fastcode]} {
# Fast methods
method bark {} {...}
method wagtail {} {...}
} else {
# Slow methods
method bark {} {...}
method wagtail {} {...}
}
}
Use a macro. For example, your snit::widget's -background option should be propagated to a number of component widgets. You could implement that like this:
snit::widget mywidget {
option -background -default white -configuremethod PropagateBackground
method PropagateBackground {option value} {
$comp1 configure $option $value
$comp2 configure $option $value
$comp3 configure $option $value
}
}
For one option, this is fine; if you've got a number of options, it becomes tedious and error prone. So package it as a macro:
snit::macro propagate {option "to" components} {
option $option -configuremethod Propagate$option
set body "\n"
foreach comp $components {
append body "\$$comp configure $option \$value\n"
}
method Propagate$option {option value} $body
}
Then you can use it like this:
snit::widget mywidget {
option -background default -white
option -foreground default -black
propagate -background to {comp1 comp2 comp3}
propagate -foreground to {comp1 comp2 comp3}
}
Yes, there are. You can't redefine any standard Tcl commands or Snit type definition statements. You can use any other command name, including the name of a previously defined macro.
If you're using Snit macros in your application, go ahead and name them in the global namespace, as shown above. But if you're using them to define types or widgets for use by others, you should define your macros in the same namespace as your types or widgets. That way, they won't conflict with other people's macros.
If my fancy snit::widget is called ::mylib::mywidget, for example, then I should define my propagate macro as ::mylib::propagate:
snit::macro mylib::propagate {option "to" components} { ... }
snit::widget ::mylib::mywidget {
option -background default -white
option -foreground default -black
mylib::propagate -background to {comp1 comp2 comp3}
mylib::propagate -foreground to {comp1 comp2 comp3}
}
This document, and the package it describes, will undoubtedly contain bugs and other problems. Please report such in the category snit of the Tcllib SF Trackers [http://sourceforge.net/tracker/?group_id=12883]. Please also report any ideas for enhancements you may have for either package and/or documentation.
BWidget, C++, Incr Tcl, adaptors, class, mega widget, object, object oriented, widget, widget adaptors
Programming tools
Copyright (c) 2003-2006, by William H. Duquette
| 2.2 | snit |